This is a guest post by Eugene Stern.
Now that I have kids in school, I’ve become a lot more familiar with high-stakes testing, which is the practice of administering standardized tests with major consequences for students who take them (you have to pass to graduate), their teachers (who are often evaluated based on standarized test results), and their school districts (state funding depends on test results). To my great chagrin, New Jersey, where I live, is in the process of putting such a teacher evaluation system in place (for a lot more detail and criticism, see here).
The excellent John Ewing pointed me to a pretty comprehensive survey of standardized testing called “Measuring Up,” by Harvard Ed School prof Daniel Koretz, who teaches a course there about this stuff. If you have any interest in the subject, the book is very much worth your time. But in case you don’t get to it, or just to whet your appetite, here are my top 10 takeaways:
Believe it or not, most of the people who write standardized tests aren’t idiots. Building effective tests is a difficult measurement problem! Koretz makes an analogy to political polling, which is a good reminder that a test result is really a sample from a distribution (if you take multiple versions of a test designed to measure the same thing, you won’t do exactly the same each time), and not an absolute measure of what someone knows. It’s also a good reminder that the way questions are phrased can matter a great deal.
The reliability of a test is inversely related to the standard deviation of this distribution: a test is reliable if your score on it wouldn’t vary very much from one instance to the next. That’s a function of both the test itself and the circumstances under which people take it. More reliability is better, but the big trade-off is that increasing the sophistication of the test tends to decrease reliability. For example, tests with free form answers can test for a broader range of skills than multiple choice, but they introduce variability across graders, and even the same person may grade the same test differently before and after lunch. More sophisticated tasks also take longer to do (imagine a lab experiment as part of a test), which means fewer questions on the test and a smaller cross-section of topics being sampled, again meaning more noise and less reliability.
A complementary issue is bias, which is roughly about people doing better or worse on a test for systematic reasons outside the domain being tested. Again, there are trade-offs: the more sophisticated the test, the more extraneous skills beyond those being tested it may be bringing in. One common way to weed out such questions is to look at how people who score the same on the overall test do on each particular question: if you get variability you didn’t expect, that may be a sign of bias. It’s harder to do this for more sophisticated tests, where each question is a bigger chunk of the overall test. It’s also harder if the bias is systematic across the test.
Beyond the (theoretical) distribution from which a single student’s score is a sample, there’s also the (likely more familiar) distribution of scores across students. This depends both on the test and on the population taking it. For example, for many years, students on the eastern side of the US were more likely to take the SAT than those in the west, where only students applying to very selective eastern colleges took the test. Consequently, the score distributions were very different in the east and the west (and average scores tended to be higher in the west), but this didn’t mean that there was bias or that schools in the west were better.
The shape of the score distribution across students carries important information about the test. If a test is relatively easy for the students taking it, scores will be clustered to the right of the distribution, while if it’s hard, scores will be clustered to the left. This matters when you’re interpreting results: the first test is worse at discriminating among stronger students and better at discriminating among weaker ones, while the second is the reverse.
The score distribution across students is an important tool in communicating results (you may not know right away what a score of 600 on a particular test means, but if you hear it’s one standard deviation above a mean of 500, that’s a decent start). It’s also important for calibrating tests so that the results are comparable from year to year. In general, you want a test to have similar means and variances from one year to the next, but this raises the question of how to handle year-to-year improvement. This is particularly significant when educational goals are expressed in terms of raising standardized test scores.
If you think in terms of the statistics of test score distributions, you realize that many of those goals of raising scores quickly are deluded. Koretz has a good phrase for this: the myth of the vanishing variance. The key observation is that test score distributions are very wide, on all tests, everywhere, including countries that we think have much better education systems than we do. The goals we set for student score improvement (typically, a high fraction of all students taking a test several years from now are supposed to score above some threshold) imply a great deal of compression at the lower end of this distribution – compression that has never been seen in any country, anywhere. It sounds good to say that every kid who takes a certain test in four years will score as proficient, but that corresponds to a score distribution with much less variance than you’ll ever see. Maybe we should stop lying to ourselves?
Koretz is highly critical of the recent trend to report test results in terms of standards (e.g., how many students score as “proficient”) instead of comparisons (e.g., your score is in the top 20% of all students who took the test). Standards and standard-based reporting are popular because it’s believed that American students’ performance as a group is inadequate. The idea is that being near the top doesn’t mean much if the comparison group is weak, so instead we should focus on making sure every student meets an absolute standard needed for success in life. There are three (at least) problems with this. First, how do you set a standard – i.e., what does proficient mean, anyway? Koretz gives enough detail here to make it clear how arbitrary the standards are. Second, you lose information: in the US, standards are typically expressed in terms of just four bins (advanced, proficient, partially proficient, basic), and variation inside the bins is ignored. Third, even standards-based reporting tends to slide back into comparisons: since we don’t know exactly what proficient means, we’re happiest when our school, or district, or state places ahead of others in the fraction of students classified as proficient.
Koretz’s other big theme is score inflation for high-stakes tests: if everyone is evaluated based on test scores, everyone has an incentive to get those scores up, whether or not that actually has much correlation with learning. If you remember anything from the book or from this post, remember this phrase: sawtooth pattern. The idea is that when a new high-stakes standardized test appears, average scores start at some base level, go up quickly as people figure out how to game the test, then plateau. If the test is replaced with another, the same thing happens: base, rapid growth, plateau. Repeat ad infinitum. Koretz and his collaborators did a nice experiment in which they went back to a school district in which one high-stakes test had been replaced with another and administered the first test several years later. Now that teachers weren’t teaching to the first test, scores on it reverted back to the original base level. Moral: score inflation is real, pervasive, and unavoidable, unless we bite the bullet and do away with high-stakes tests.
While Koretz is sympathetic toward test designers, who live the complexity of standardized testing every day, he is harsh on those who (a) interpret and report on test results and (b) set testing and education policy, without taking that complexity into account. Which, as he makes clear, is pretty much everyone who reports on results and sets policy.
If you think it’s a good idea to make high-stakes decisions about schools and teachers based on standardized test results, Koretz’s book offers several clear warnings.
First, we should expect any high-stakes test to be gamed. Worse yet, the more reliable tests, being more predictable, are probably easier to game (look at the SAT prep industry).
Second, the more (statistically) reliable tests, by their controlled nature, cover only a limited sample of the domain we want students to learn. Tests trying to cover more ground in more depth (“tests worth teaching to,” in the parlance of the last decade) will necessarily have noisier results. This noise is a huge deal when you realize that high-stakes decisions about teachers are made based on just two or three years of test scores.
Third, a test that aims to distinguish “proficiency” will do a worse job of distinguishing students elsewhere in the skills range, and may be largely irrelevant for teachers whose students are far away from the proficiency cut-off. (For a truly distressing example of this, see here.)
With so many obstacles to rating schools and teachers reliably based on standardized test scores, is it any surprise that we see results like this?
My friend Josh Vekhter sent me this blog post written by someone who calls herself celandine13 and tutors students with learning disabilities.
In the post, she reframes the concept of mistake or “being bad at something” as often stemming from some fundamental misunderstanding or poor procedure:
Once you move it to “you’re performing badly because you have the wrong fingerings,” or “you’re performing badly because you don’t understand what a limit is,” it’s no longer a vague personal failing but a causal necessity. Anyone who never understood limits will flunk calculus. It’s not you, it’s the bug.
This also applies to “lazy.” Lazy just means “you’re not meeting your obligations and I don’t know why.” If it turns out that you’ve been missing appointments because you don’t keep a calendar, then you’re not intrinsically “lazy,” you were just executing the wrong procedure. And suddenly you stop wanting to call the person “lazy” when it makes more sense to say they need organizational tools.
And she wants us to stop with the labeling and get on with the understanding of why the mistake was made and addressing that, like she does when she tutors students. She even singles out certain approaches she considers to be flawed from the start:
This is part of why I think tools like Knewton, while they can be more effective than typical classroom instruction, aren’t the whole story. The data they gather (at least so far) is statistical: how many questions did you get right, in which subjects, with what learning curve over time? That’s important. It allows them to do things that classroom teachers can’t always do, like estimate when it’s optimal to review old material to minimize forgetting. But it’s still designed on the error model. It’s not approaching the most important job of teachers, which is to figure out why you’re getting things wrong — what conceptual misunderstanding, or what bad study habit, is behind your problems. (Sometimes that can be a very hard and interesting problem. For example: one teacher over many years figured out that the grammar of Black English was causing her students to make conceptual errors in math.)
On the one hand I like the reframing: it’s always good to see knee-jerk reactions become more contemplative, and it’s always good to see people trying to help rather than trying to blame. In fact, one of my tenets of real life is that mistakes will be made, and it’s not the mistake that we should be anxious about but how we act to fix the mistake that exposes who we are as people.
I would, however, like to take issue with her anti-example in the case of Knewton, which is an online adaptive learning company. Full disclosure: I interviewed with Knewton before I took my current job, and I like the guys who work there. But, I’d add, I like them partly because of the healthy degree of skepticism they take with them to their jobs.
What the blogwriter celandine13 is pointing out, correctly, is that understanding causality is pretty awesome when you can do it. If you can figure out why someone is having trouble learning something, and if you can address that underlying issue, then fixing the consequences of that issue get a ton easier. Agreed, but I have three points to make:
- First, a non-causal data mining engine such as Knewton will also stumble upon a way to fix the underlying problem by dint of having a ton of data and noting that people who failed a calculus test, say, did much better after having limits explained to them in a certain way. This is much like the spellcheck engine of Google works by keeping track of previous spelling errors, and not by mind reading how people think about spelling wrong.
- Second, it’s not always easy to find the underlying cause of bad testing performance, even if you’re looking for it directly. I’m not saying it’s fruitless – tutors I know are incredibly good at that – but there’s room for both “causality detectives” and tons of smart data mining in this field.
- Third, it’s definitely not always easy to address the underlying cause of bad test performance. If you find out that the grammar of Black English affects students’ math test scores, what do you do about it?
Having said all that, I’d like to once more agree with the underlying message that a mistake is a first and foremost a signal rather than a reflection of someone’s internal thought processes. The more we think of mistakes as learning opportunities the faster we learn.
When I worked as a research mathematician, I was always flabbergasted by the speed at which other people would seem to absorb mathematical theory. I had then, and pretty much have now, this inability to believe anything that I can’t prove from first principles, or at least from stuff I already feel completely comfortable with. For me, it’s essentially mathematically unethical to use a result I can’t prove or at least understand locally.
I only recently realized that not everyone feels this way. Duh. People often just assemble accepted facts about a field quickly just to explore the landscape and get the feel for something – it makes complete sense to me now that one can do this and it doesn’t seem at all weird. And it explains what I saw happening in grad school really well too.
Most people just use stuff they “know to be true,” without having themselves gone through the proof. After all, things like Deligne’s work on Weil Conjectures or Gabber’s recent work on finiteness of etale cohomology for pseudo-excellent schemes are really fucking hard, and it’s much more efficient to take their results and use them than it is to go through all the details personally.
After all, I use a microwave every day without knowing how it works, right?
I’m not sure I know where I got the feeling that this was an ethical issue. Probably it happened without intentional thought, when I was learning what a proof is in math camp, and I’d perhaps state a result and someone would say, how do you know that? and I’d feel like an asshole unless I could prove it on the spot.
Anyway, enough about me and my confused definition of mathematical ethics – what I now realize is that, as mathematics is developed more and more, it will become increasingly difficult for a graduate student to learn enough and then prove an original result without taking things on faith more and more. The amount of mathematical development in the past 50 years is just frighteningly enormous, especially in certain fields, and it’s just crazy to imagine someone learning all this stuff in 2 or 3 years before working on a thesis problem.
What I’m saying, in other words, is that my ethical standards are almost provably unworkable in modern mathematical research. Which is not to say that, over time, a person in a given field shouldn’t eventually work out all the details to all the things they’re relying on, but it can’t be linear like I forced myself to work.
And there’s a risk, too: namely, that as people start getting used to assuming hard things work, fewer mistakes will be discovered. It’s a slippery slope.
I’m happy to say that the book I’m writing with Rachel Schutt called Doing Data Science is officially out for early review. That means a few chapters which we’ve deemed “ready” have been sent to some prominent people in the field to see what they think. Thanks, prominent and busy people!
It also means that things are (knock on wood) wrapping up on the editing side. I’m cautiously optimistic that this book will be a valuable resource for people interested in what data scientists do, especially people interested in switching fields. The range of topics is broad, which I guess means that the most obvious complaint about the book will be that we didn’t cover things deeply enough, and perhaps that the level of pre-requisite assumptions is uneven. It’s hard to avoid.
Thanks to my awesome editor Courtney Nash over at O’Reilly for all her help!
And by the way, we have an armadillo on our cover, which is just plain cool:
This is a guest post by Eugene Stern.
A big reason I love this blog is Cathy’s war on crappy models. She has posted multiple times already about the lousy performance of models that rate teachers based on year-to-year changes in student test scores (for example, read about it here). Much of the discussion focuses on the model used in New York City, but such systems have been, or are being, put in place all over the country. I want to let you know about the version now being considered for use across the river, in New Jersey. Once you’ve heard more, I hope you’ll help me try to stop it.
A little background if you haven’t heard about this before. Because it makes no sense to rate teachers based on students’ absolute grades or test scores (not all students start at the same place each year), the models all compare students’ test scores against some baseline. The simplest thing to do is to compare each student’s score on a test given at the end of the school year against their score on a test given at the end of the previous year. Teachers are then rated based on how much their students’ scores improved over the year.
Comparing with the previous year’s score controls for the level at which students start each year, but not for other factors beside the teacher that affect how much they learn. This includes attendance, in-school environment (curriculum, facilities, other students in the class), out-of-school learning (tutoring, enrichment programs, quantity and quality of time spent with parents/caregivers), and potentially much more. Fancier models try to take these into account by comparing each student’s end of year score with a predicted score. The predicted score is based both on the student’s previous score and on factors like those above. Improvement beyond the predicted score is then attributed to the teacher as “value added” (hence the name “value-added models,” or VAM) and turned into a teacher rating in some way, often using percentiles. One such model is used to rate teachers in New York City.
It’s important to understand that there is no single value-added model, rather a family of them, and that the devil is in the details. Two different teacher rating systems, based on two models of the predicted score, may perform very differently – both across the board, and in specific locations. Different factors may be more or less important depending on where you are. For example, income differences may matter more in a district that provides few basic services, so parents have to pay to get extracurriculars for their kids. And of course the test itself matters hugely as well.
Testing the VAM models
Teacher rating models based on standardized tests have been around for 25 years or so, but two things have happened in the last decade:
- Some people started to use the models in formal teacher evaluation, including tenure decisions.
- Some (other) people started to test the models.
This did not happen in the order that one would normally like. Wanting to make “data-driven decisions,” many cities and states decided to start rating teachers based on “data” before collecting any data to validate whether that “data” was any good. This is a bit like building a theoretical model of how cancer cells behave, synthesizing a cancer drug in the lab based on the model, distributing that drug widely without any trials, then waiting around to see how many people die from the side effects.
The full body count isn’t in yet, but the models don’t appear to be doing well so far. To look at some analysis of VAM data in New York City, start here and here. Note: this analysis was not done by the city but by individuals who downloaded the data after the city had to make it available because of disclosure laws.
I’m not aware of any study on the validity of NYC’s VAM ratings done by anyone actually affiliated with the city – if you know of any, please tell me. Again, the people preaching data don’t seem willing to actually use data to evaluate the quality of the systems they’re putting in place.
Assuming you have more respect for data than the mucky-mucks, let’s talk about how well the models actually do. Broadly, two ways a model can fail are being biased and being noisy. The point of the fancier value-added models is to try to eliminate bias by factoring in everything other than the teacher that might affect a student’s test score. The trouble is that any serious attempt to do this introduces a bunch of noise into the model, to the degree that the ratings coming out look almost random.
You’d think that a teacher doesn’t go from awful to great or vice versa in one year, but the NYC VAM ratings show next to no correlation in a teacher’s rating from one year to the next. You’d think that a teacher either teaches math well or doesn’t, but the NYC VAM ratings show next to no correlation in a teacher’s rating teaching a subject to one grade and their rating teaching it to another – in the very same year! (Gary Rubinstein’s blog, linked above, documents these examples, and a number of others.) Again, this is one particular implementation of a general class of models, but using such noisy data to make significant decisions about teachers’ careers seems nuts.
What’s happening in New Jersey
With all this as background, let’s turn to what’s happening in New Jersey.
You may be surprised that the version of the model proposed by Chris Christie‘s administration (the education commissioner is Christie appointee Chris Cerf, who helped put VAM in place in NYC) is about the simplest possible. There is no attempt to factor out bias by trying to model predicted scores, just a straight comparison between this year’s standardized test score and last year’s. For an overview, see this.
In more detail, the model groups together all students with the same score on last year’s test, and represents each student’s progress by their score on this year’s test, viewed as a percentile across this group. That’s it. A fancier version uses percentiles calculated across all students with the same score in each of the last several years. These can’t be calculated explicitly (you may not find enough students that got exactly the same score each the last few years), so they are estimated, using a statistical technique called quantile regression.
By design, both the simple and the fancy version ignore everything about a student except their test scores. As a modeler, or just as a human being, you might find it silly not to distinguish between a fourth grader in a wealthy suburb who scored 600 on a standardized test from a fourth grader in the projects with the same score. At least, I don’t know where to find a modeler who doesn’t find it silly, because nobody has bothered to study the validity of using this model to rate teachers. If I’m wrong, please point me to a study.
Politics and SGP
But here we get into the shell game of politics, where rating teachers based on the model is exactly the proposal that lies at the end of an impressive trail of doubletalk. Follow the bouncing ball.
These models, we are told, differ fundamentally from VAM (which is now seen as somewhat damaged goods politically, I suspect). While VAM tried to isolate teacher contribution, these models do no such thing – they are simply measuring student progress from year to year, which, after all, is what we truly care about. The models have even been rebranded with a new name: student growth percentiles, or SGP. SGP is sold as just describing student progress rather than attributing it to teachers, there can’t be any harm in that, right? – and nothing that needs validation, either. And because SGP is such a clean methodology – if you’re looking for a data-driven model to use for broad “educational assessment,” don’t get yourself into that whole VAM morass, use SGP instead!
Only before you know it, educational assessment turns into, you guessed it, rating teachers. That’s right: because these models aren’t built to rate teachers, they can focus on the things that really matter (student progress), and thus end up being – wait for it – much better for rating teachers! War is peace, friends. Ignorance is strength.
Creators of SGP
You can find a good discussion of SGP’s and their use in evaluation here, and a lot more from the same author, the impressively prolific Bruce Baker, here. Here’s a response from the creators of SGP. They maintain that information about student growth is useful (duh), and agree that differences in SGP’s should not be attributed to teachers (emphasis mine):
Large-scale assessment results are an important piece of evidence but are not sufficient to make causal claims about school or teacher quality.
SGP and teacher evaluations
But guess what?
The New Jersey Board of Ed and state education commissioner Cerf are putting in place a new teacher evaluation code, to be used this coming academic year and beyond. You can find more details here and here.
Summarizing: for math and English teachers in grades 4-8, 30% of their annual evaluation next year would be mandated by the state to come from those very same SGP’s that, according to their creators, are not sufficient to make causal claims about teacher quality. These evaluations are the primary input in tenure decisions, and can also be used to take away tenure from teachers who receive low ratings.
The proposal is not final, but is fairly far along in the regulatory approval process, and would become final in the next several months. In a recent step in the approval process, the weight given to SGP’s in the overall evaluation was reduced by 5%, from 35%. However, the 30% weight applies next year only, and in the future the state could increase the weight to as high as 50%, at its discretion.
Modeler’s Note #1: the precise weight doesn’t really matter. If the SGP scores vary a lot, and the other components don’t vary very much, SGP scores will drive the evaluation no matter what their weight.
Modeler’s Note #2: just reminding you again that this data-driven framework for teacher evaluation is being put in place without any data-driven evaluation of its effectiveness. And that this is a feature, not a bug – SGP has not been tested as an attribution tool because we keep hearing that it’s not meant to be one.
In a slightly ironic twist, commissioner Cerf has responded to criticisms that SGP hasn’t been tested by pointing to a Gates Foundation study of the effectiveness of… value-added models. The study is here. It draws pretty positive conclusions about how well VAM’s work. A number of critics have argued, pretty effectively, that the conclusions are unsupported by the data underlying the study, and that the data actually shows that VAM’s work badly. For a sample, see this. For another example of a VAM-positive study that doesn’t seem to stand up to scrutiny, see this and this.
Modeler’s Role Play #1
Say you were the modeler who had popularized SGP’s. You’ve said that the framework isn’t meant to make causal claims, then you see New Jersey (and other states too, I believe) putting a teaching evaluation model in place that uses SGP to make causal claims, without testing it first in any way. What would you do?
So far, the SGP mavens who told us that “Large-scale assessment results are an important piece of evidence but are not sufficient to make causal claims about school or teacher quality” remain silent about the New Jersey initiative, as far as I know.
Modeler’s Role Play #2
Now you’re you again, and you’ve never heard about SGP’s and New Jersey’s new teacher evaluation code until today. What do you do?
I want you to help me stop this thing. It’s not in place yet, and I hope there’s still time.
I don’t think we can convince the state education department on the merits. They’ve made the call that the new evaluation system is better than the current one or any alternatives they can think of, they’re invested in that decision, and we won’t change their minds directly. But we can make it easier for them to say no than to say yes. They can be influenced – by local school administrators, state politicians, the national education community, activists, you tell me who else. And many of those people will have more open minds. If I tell you, and you tell the right people, and they tell the right people, the chain gets to the decision makers eventually.
I don’t think I could convince Chris Christie, but maybe I could convince Bruce Springsteen if I met him, and maybe Bruce Springsteen could convince Chris Christie.
I thought we could start with a manifesto – a direct statement from the modeling community explaining why this sucks. Directed at people who can influence the politics, and signed by enough experts (let’s get some big names in there) to carry some weight with those influencers.
Can you help? Help write it, sign it, help get other people to sign it, help get it to the right audience. Know someone whose opinion matters in New Jersey? Then let me know, and help spread the word to them. Use Facebook and Twitter if it’ll help. And don’t forget good old email, phone calls, and lunches with friends.
Or, do you have a better idea? Then put it down. Here. The comments section is wide open. Let’s not fall back on criticizing the politicians for being dumb after the fact. Let’s do everything we can to keep them from doing this dumb thing in the first place.
Shame on us if we can’t make this right.
This is a guest post by Kaisa Taipale. Kaisa got a BS at Caltech, a Ph.D. in math at the University of Minnesota, was a post-doc at MSRI, an assistant professor at St. Olaf College 2010-2012, and is currently visiting Cornell, which is where I met here a couple of weeks ago, and where she told me about her cool visualizations of math Ph.D. emigration patterns and convinced her to write a guest post. Here’s Kaisa on a bridge:
Math data and viz
I was inspired by this older post on Mathbabe, about visualizing the arXiv postings of various math departments.
It got me thinking about tons of interesting questions I’ve asked myself and could answer with visualizations: over time, what’s been coolest on the arXiv? are there any topics that are especially attractive to hiring institutions? There’s tons of work to do!
I had to start somewhere though, and as I’m a total newbie when it comes to data analysis, I decided to learn some skills while focusing on a data set that I have easy non-technical access to and look forward to reading every year. I chose the AMS Annual Survey. I also wanted to stick to questions really close to my thoughts over the last two years, namely the academic job search.
I wanted to learn to use two tools, R and Circos. Why Circos? See the visualizations of college major and career path here - it’s pretty! I’ve messed around with a lot of questions, but in this post I’ll look at two and a half.
Where do graduating PhDs from R1 universities end up, in the short term? I started with graduates of public R1s, as I got my PhD at one.
The PhD-granting institutions are colored green, while academic institutions granting other degrees are in blue. Purple is for business, industry, government, and research institutions. Red is for non-U.S. employment or people not seeking — except for the bright red, which is still seeking. Yellow rounds things out at unknown. Remember, these figures are for immediate plans after graduation rather than permanent employment.
While I was playing with this data (read “learning how to use the reshape and ggplot2 packages”) I noticed that people from private R1s tend to end up at private R1s more often. So I graphed that too.
Does the professoriate in the audience have any idea if this is self-selection or some sort of preference on the part of employers? Also, what happened between 2001 and 2003? I was still in college, and have no idea what historical events are at play here.
Where mathematicians go
For any given year, we can use a circular graph to show us where people go. This is a more clumped version of the above data from 2010 alone, plotted using Circos. (Supplemental table E.4 from the AMS report online.)
The other question – the question current mathematicians secretly care more about, in a gossipy and potentially catty way – is what fields lead to what fate. We all know algebra and number theory are the purest and most virtuous subjects, and applied math is for people who want to make money or want to make a difference in the world.
[On that note, you might notice that I removed statistics PhDs in the visualization below, and I also removed some of the employment sectors that gained only a few people a year. The stats ribbons are huge and the small sectors are very small, so for looks alone I took them out.]
Higher resolution version available here.
I wish I could animate a series of these to show this view over time as well. Let me know if you know how to do that! Another nice thing I could do would be to set up a webpage in which these visualizations could be explored in a bit more depth. (After finals.)
- I haven’t computed any numbers for you
- the graphs from R show employment in each field by percentage of graduates instead of total number per category;
- it’s hard to show both data over time and all the data one could explore. But it’s a start.
I should finish with a shout-out to Roger Peng and Jeff Leek, though we’ve never met: I took Peng’s Computing for Data Analysis and much of Leek’s Data Analysis on Coursera (though I’m one of those who didn’t finish the class). Their courses and Stack Overflow taught me almost everything I know about R. As I mentioned above, I’m pretty new to this type of analysis.
What questions would you ask? How can I make the above cooler? Did you learn anything?
There’ve been a couple of articles in the past few days about teacher Value-Added Testing that have enraged me.
If you haven’t been paying attention, the Value-Added Model (VAM) is now being used in a majority of the states (source: the Economist):
But it gives out nearly random numbers, as gleaned from looking at the same teachers with two scores (see this previous post). There’s a 24% correlation between the two numbers. Note that some people are awesome with respect to one score and complete shit on the other score:
Final thing you need to know about the model: nobody really understands how it works. It relies on error terms of an error-riddled model. It’s opaque, and no teacher can have their score explained to them in Plain English.
Now, with that background, let’s look into these articles.
First, there’s this New York Times article from yesterday, entitled “Curious Grade for Teachers: Nearly All Pass”. In this article, it describes how teachers are nowadays being judged using a (usually) 50/50 combination of classroom observations and VAM scores. This is different from the past, which was only based on classroom observations.
What they’ve found is that the percentage of teachers found “effective or better” has stayed high in spite of the new system – the numbers are all over the place but typically between 90 and 99 percent of teachers. In other words, the number of teachers that are fingered as truly terrible hasn’t gone up too much. What a fucking disaster, at least according to the NYTimes, which seems to go out of its way to make its readers understand how very much high school teachers suck.
A few things to say about this.
- Given that the VAM is nearly a random number generator, this is good news – it means they are not trusting the VAM scores blindly. Of course, it still doesn’t mean that the right teachers are getting fired, since half of the score is random.
- Another point the article mentions is that failing teachers are leaving before the reports come out. We don’t actually know how many teachers are affected by these scores.
- Anyway, what is the right number of teachers to fire each year, New York Times? And how did you choose that number? Oh wait, you quoted someone from the Brookings Institute: “It would be an unusual profession that at least 5 percent are not deemed ineffective.” Way to explain things so scientifically! It’s refreshing to know exactly how the army of McKinsey alums approach education reform.
- The overall article gives us the impression that if we were really going to do our job and “be tough on bad teachers,” then we’d weight the Value-Added Model way more. But instead we’re being pussies. Wonder what would happen if we weren’t pussies?
The second article explained just that. It also came from the New York Times (h/t Suresh Naidu), and it was a the story of a School Chief in Atlanta who took the VAM scores very very seriously.
What happened next? The teachers cheated wildly, changing the answers on their students’ tests. There was a big cover-up, lots of nasty political pressure, and a lot of good people feeling really bad, blah blah blah. But maybe we can take a step back and think about why this might have happened. Can we do that, New York Times? Maybe it had to do with the $500,000 in “performance bonuses” that the School Chief got for such awesome scores?
Let’s face it, this cheating scandal, and others like it (which may never come to light), was not hard to predict (as I explain in this post). In fact, as a predictive modeler, I’d argue that this cheating problem is the easiest thing to predict about the VAM, considering how it’s being used as an opaque mathematical weapon.
MathBabe recently wrote an article critical of the elitist nature of Ted Talks, which you can read here. Fortunately for her, and for the hoi polloi everywhere clamoring for populist science edutainment, there is an alternative: Nerd Nite. Once a month, in cities all over the globe, nerds herd into a local bar and turn it into a low-brow forum for innovative science ideas. Think Ted Talks on tequila.
Each month, three speakers present talks for 20-30 minutes, followed by questions and answers from the invariably sold-out audience. The monthly forum gives professional and amateur scientists an opportunity to explain their fairly abstruse specialties accessibly to a lay audience – a valuable skill. Since the emphasis is on science entertainment, it also gives the speakers a chance to present their ideas in a more engaging way: in iambic pentameter, in drag with a tuba, in three-part harmony, or via interpretive dance – an invaluable skill. The resulting atmosphere is informal, delightfully debauched, and refreshingly pro-science.
Slaking our thirst for both science education and mojitos, Nerd Nite started small but quickly went viral. Nerd Nites are now being held in 50 cities, from San Francisco to Kansas City and Auckland to Liberia. You can find the full listing of cities here; if you don’t see one near you, start one!
Last Wednesday night I was twitterpated to be one of three guest nerds sharing the stage at San Francisco’s Nerd Nite. I put the chic back into geek with a biology talk entitled “Genital Plugs, Projectile Penises, and Gay Butterflies: A Naturalist Explains the Birds and the Bees.”
A video recording of the presentation will be available online soon, but in the meantime, here’s a tantalizing clip from the talk, in which Isabella Rossellini explains the mating habits of the bee. Warning: this is scientifically sexy.
I shared the stage with Chris Anderson, who gave a fascinating talk on how the DIY community is building drones out of legos and open-source software. These DIY drones fly below government regulation and can be used for non-military applications, something we hear far too little of in the daily war digest that passes for news. The other speaker was Mark Rosin of the UK-based Guerrilla Science project. This clever organization reaches out to audiences at non-science venues, such as music concerts, and conducts entertaining presentations that teach core science ideas. As part of his presentation Mark used 250 inflated balloons and a bass amp to demonstrate the physics concept of resonance.
If your curiosity has been piqued and you’d like to check out an upcoming Nerd Nite, consider attending the upcoming Nerdtacular, the first Nerd Nite Global Festival, to be held this August 16-18th in Brooklyn, New York.
The global Nerdtacular: Now that’s an idea worth spreading.
At my new job I’ve been spending my time editing my book with Rachel Schutt (who is joining me at JRL next week! Woohoo!). It’s called Doing Data Science and it’s based on these notes I took when she taught a class on data science at Columbia last semester. Right now I’m working on the alternating least squares chapter, where we learned from Matt Gattis how to build and optimize a recommendation system. A very cool algorithm.
However, to be honest I’ve started to feel very sorry for the one parameter we call It’s also sometimes referred to as “the prior”.
Let me tell you, the world is asking too much from this little guy, and moreover most of the big-data world is too indifferent to its plight. Let me explain.
First, he’s supposed to reflect an actual prior belief – namely, his size is supposed to reflect a mathematical vision of how big we think the coefficients in our solution should be.
In an ideal world, we would think deeply about this question of size before looking at our training data, and think only about the scale of our data (i.e. the input), the scale of the preferences (i.e. the recommendation system output) and the quality and amount of training data we have, and using all of that, we’d figure out our prior belief on the size or at least the scale of our hoped-for solution.
I’m not statistician, but that’s how I imagine I’d spend my days if I were: thinking through this reasoning carefully, and even writing it down carefully, before I ever start my training. It’s a discipline like any other to carefully state your beliefs beforehand so you know you’re not just saying what the data wants to hear.
as convergence insurance
But then there’s the next thing we ask of our parameter namely we assign him the responsibility to make sure our algorithm converges.
Because our algorithm isn’t a closed form solution, but rather we are discovering coefficients of two separate matrices and , fixing one while we tweak the other, then switching. The algorithm stops when, after a full cycle of fixing and tweaking, none of the coefficients have moved by more than some pre-ordained
The fact that this algorithm will in fact stop is not obvious, and in fact it isn’t always true.
It is (mostly*) true, however, if our little is large enough, which is due to the fact that our above-mentioned imposed belief of size translates into a penalty term, which we minimize along with the actual error term. This little miracle of translation is explained in this post.
And people say that all the time. When you say, “hey what if that algorithm doesn’t converge?” They say, “oh if is big enough it always does.”
But that’s kind of like worrying about your teenage daughter getting pregnant so you lock her up in her room all the time. You’ve solved the immediate problem by sacrificing an even bigger goal.
Because let’s face it, if the prior is too big, then we are sacrificing our actual solution for the sake of conveniently small coefficients and convergence. In the asymptotic limit, which I love thinking about, our coefficients all go to zero and we get nothing at all. Our teenage daughter has run away from home with her do-nothing boyfriend.
By the way, there’s a discipline here too, and I’d suggest that if the algorithm doesn’t converge you might also want to consider reducing your number of latent variables rather than increasing your since you could be asking too much from your training data. It just might not be able to distinguish that many important latent characteristics.
as tuning parameter
Finally, we have one more job for our little , we’re not done with him yet. Actually for some people this is his only real job, because in practice this is how he’s treated. Namely, we optimize him so that our results look good under whatever metric we decide to care about (but it’s probably the mean squared error of preference prediction on a test set (hopefully on a test set!)).
In other words, in reality most of the above nonsense about is completely ignored.
This is one example among many where having the ability to push a button that makes something hard seem really easy might be doing more harm than good. In this case the button says “optimize with respect to “, but there are other buttons that worry me just as much, and moreover there are lots of buttons being built right now that are even more dangerous and allow the users to be even more big-data-blithe.
I’ve said it before and I’ll say it again: you do need to know about inverting a matrix, and other math too, if you want to be a good data scientist.
* There’s a change-of-basis ambiguity that’s tough to get rid of here, since you only choose the number of latent variables, not their order. This doesn’t change the overall penalty term, so you can minimize that with large enough but if you’re incredibly unlucky I can imagine you might bounce between different solutions that differ by a base change. In this case your steps should get smaller, i.e. the amount you modify your matrix each time you go through the algorithm. This is only a theoretical problem by the way but I’m a nerd.
If this New York Times editorial is correct, and it certainly passes the smell test, students are not well-served by online courses but are by so-called “hybrid” courses, where there’s a bit of online stuff and also a bit of one-on-one time. From the editorial:
The research has shown over and over again that community college students who enroll in online courses are significantly more likely to fail or withdraw than those in traditional classes, which means that they spend hard-earned tuition dollars and get nothing in return. Worse still, low-performing students who may be just barely hanging on in traditional classes tend to fall even further behind in online courses.
This is important news for math departments, at least in the medium term (i.e. until machine learners figure out how to successfully simulate one-on-one interactions), because it means they won’t be replacing calculus class with a computer. And as every mathematician should know, calculus is the bread and butter of math departments.
I don’t agree with everything she always says, but I agree with everything Izabella Laba says in this post called Gender Bias 101 For Mathematicians (hat tip Jordan Ellenberg). And I’m kind of jealous she put it together in such a fantastic no-bullshit way.
Namely, she debunks a bunch of myths of gender bias. Here’s my summary, but you should read the whole thing:
- Myth: Sexism in math is perpetrated mainly by a bunch of enormously sexist old guys. Izabella: Nope, it’s everyone, and there’s lots of evidence for that.
- Myth: The way to combat sexism is to find those guys and isolate them. Izabella: Nope, that won’t work, since it’s everyone.
- Myth: If it’s really everyone, it’s too hard to solve. Izabella: Not necessarily, and hey you are still trying to solve the Riemann Hypothesis even though that’s hard (my favorite argument).
- Myth: We should continue to debate about its existence rather than solution. Izabella: We are beyond that, it’s a waste of time, and I’m not going to waste my time anymore.
- Myth: Izabella, you are only writing this to be reassured. Izabella: Don’t patronize me.
Here’s what I’d add. I’ve been arguing for a long time that gender bias against girls in math starts young and starts at the cultural level. It has to do with expectations of oneself just as much as a bunch of nasty old men (by the way, the above is not to say there aren’t nasty old men (and nasty old women!), just that it’s not only about them).
My argument has been that the cultural differences are larger than the talent differences, something Larry Summers strangely dismissed without actually investigating in his famous speech.
And I think I’ve found the smoking gun for my side of this argument, in the form of an interactive New York Times graphic from last week’s Science section which I’ve screenshot here:
What this shows is that 15-year-old girls out-perform 15-year-old boys in certain countries and under-perform them in others. Those countries where they outperform boys is not random and has everything to do with cultural expectations and opportunities for girls in those countries and is explained to some extent by stereotype threat. Go read the article, it’s fascinating.
I’ll say again what I said already at the end of this post: the great news is that it is possible to address stereotype threat directly, which won’t solve everything but will go a long way.
You do it by emphasizing that mathematical talent is not inherent, nor fixed at birth, and that you can cultivate it and grow it over time and through hard work. I make this speech whenever I can to young people. Spread the word!
Last night I was waiting for a bus to go hang with my Athena Mastermind group, which consists of a bunch of very cool Barnard student entrepreneurs and their would-be role models (I say would-be because, although we role models are also very cool, I often think the students are role modeling for us).
As I was waiting at the bus stop, I overheard two women talking about the new Applied Data Science class that just started at Columbia, which is being taught by Ian Langmore, Daniel Krasner and Chang She. I knew about this class because Ian came to advertise it last semester in Rachel Schutt’s Intro to Data Science class which I blogged. One of the women at the bus stop had been in Rachel’s class and the other is in Ian’s.
Turns out I just love overhearing nerd girls talking data science at the bus stop. Don’t you??
And to top off the nerd girl experience, I’m on my way today to Nebraska to give a talk to a bunch of undergraduate women in math about what they can do with math outside of academia. I’m planning it to be an informative talk, but that’s really just cover to its real goal, which is to give a pep talk.
My experience talking to young women in math, at least when they are grad students, is that they respond viscerally to encouragement, even if it’s vague. I can actually see their egos inflate in the audience as I speak, and that’s a good thing, that’s why I’m there.
As a community, I’ve realized, nerd girls going through grad school are virtually starved for positive feedback, and so my job is pretty clear cut: I’m going to tell them how awesome they are and answer their questions about what it’s like in the “real world” and then go back to telling them how awesome they are.
By the end they sit a bit straighter and smile a bit more after I’m done, after I’ve told them, or reminded them at least, how much power they have as nerd girls – how many options they have, and how they don’t have to be risk-averse, and how they never need to apologize.
Tomorrow my audience is undergraduates, which is a bit trickier, since as an undergrad you still get consistent feedback in the form of grades. So I will tailor my information as well as my encouragement a bit, and try not to make grad school sound too scary, because I do think that getting a Ph.D. is still a huge deal. Comment below if you have suggestions for my talk, please!
I had a great time giving my “Weapons of Math Destruction” talk in San Diego, and the audience was fantastic and thoughtful.
One question that someone asked was whether the US News & World Reports college ranking model should be forced to be open sourced – wouldn’t that just cause colleges to game the model?
First of all, colleges are already widely gaming the model and have been for some time. And that gaming is a distraction and has been heading colleges in directions away from good instruction, which is a shame.
And if you suggest that they change the model all the time to prevent this, then you’ve got an internal model of this model that needs adjustment. They might be tinkering at the edges but overall it’s quite clear what’s going into the model: namely, graduation rates, SAT scores, number of Ph.D’s on staff, and so on. The exact percentages change over time but not by much.
The impact that this model has had on education and how universities apportion resources has been profound. Academic papers have been written on the law school version of this story.
Moreover, the tactics that US News & World Reports uses to enforce their dominance of the market are bullying, as you can learn from the President of Reed College, which refuses to be involved.
Back to the question. Just as I realize that opening up all data is not reasonable or desirable, because first of all there are serious privacy issues but second of all certain groups have natural advantages to openly shared resources, it is also true that opening up all models is similarly problematic.
However, certain data should surely be open: for example, the laws of our country, that we are all responsible to know, should be freely available to us (something that Aaron Swartz understood and worked towards). How can we be held responsible for laws we can’t read?
Similarly, public-facing models, such as credit scoring models and teacher value-added models, should absolutely be open and accessible to the public. If I’m being judged and measured and held accountable by some model in my daily life as a citizen, that has real impact on how my future will unfold, then I should know how that process works.
And if you complain about the potential gaming of those public-facing models, I’d answer: if they are gameable then they shouldn’t be used, considering the impact they have on so many people’s lives. Because a gameable model is a weak model, with proxies that fail.
Another way to say this is we should want someone to “game” the credit score model if it means they pay their bills on time every month (I wrote about this here).
Back to the US News & World Report model. Is it public facing? I’m no lawyer but I think a case can be made that it is, and that the public’s trust in this model makes it a very important model indeed. Evidence can be gathered by measuring the extent to which colleges game the model, which they only do because the public cares so much about the rankings.
Even so, what difference would that make, to open it up?
In an ideal world, where the public is somewhat savvy about what models can and cannot do, opening up the US News & World Reports college ranking model would result in people losing faith in it. They’d realize that it’s no more valuable than an opinion from a highly vocal uncle of theirs who is obsessed with certain metrics and blind to individual eccentricities and curriculums that may be a perfect match for a non-conformist student. It’s only one opinion among many, and not to be religiously believed.
But this isn’t an ideal world, and we have a lot of work to do to get people to understand models as opinions in this sense, and to get people to stop trusting them just because they’re mathematically presented.
I’m giving a talk at the Joint Mathematics Meeting on Thursday (it’s a 30 minute talk that starts at 11:20am, in Room 2 of the Upper Level of the San Diego Conference Center, I hope you come!).
Thinking about that talk brought something up for me that I think I want to address before the next talk. Namely, at the beginning of the talk I was explaining the title, “How Mathematics is Used Outside of Academia,” and I mentioned that most mathematicians that leave academia end up doing modeling.
I can’t remember the exact exchange, but I referred to myself at some point in this discussion as a mathematician outside of academia, at which point someone in the audience expressed incredulity:
him: Really? Are you still a mathematician? Do you prove theorems?
me: No, I don’t prove theorems any longer, now that I am a modeler… (confused look)
At the moment I didn’t have a good response to this, because he was using a different definition of “mathematician” than I was. For some reason he thought a mathematician must prove theorems.
I don’t think so. I had a conversation about this after my talk with Bob Beals, who was in the audience and who taught many years ago at the math summer program I did last summer. After getting his Ph.D. in math, Bob worked for the spooks, and now he works for RenTech. So he knows a lot about doing math outside academia too, and I liked his perspective on this question.
Namely, he wanted to look at the question through the lens of “grunt work”, which is to say all of the actual work that goes into a “result.”
As a mathematician, of course, you don’t simply sit around all day proving theorems. Actually you spend most of your time working through examples to get a feel for the terrain, and thinking up simple ways to do what seems like hard things, and trying out ideas that fail, and going down paths that are dry. If you’re lucky, then at the end of a long journey like this, you will have a theorem.
The same basic thing happens in modeling. You spend lots of time with the data, getting to know it, and then trying out certain approaches, which sometimes, or often, end up giving you nothing interesting, and half the time you realize you were expecting the wrong thing so you have to change it entirely. In the end you may end up with a model which is useful. If you’re lucky.
There’s a lot of grunt work in both endeavors, and there’s a lot of hard thinking along the way, lots of ways for you to fool yourself that you’ve got something when you haven’t. Perhaps in modeling it’s easier to lie, which is a big difference indeed. But if you’re an honest modeler then I claim the difference in the process of getting an interesting and important result is not that different.
And, I claim, I am still being a mathematician while I’m doing it.
I lied yesterday, as a friend at my Occupy meeting pointed out to me last night.
I made it seem like I look into every model before trusting it, and of course that’s not true. I eat food grown and prepared by other people daily. I go on airplanes and buses all the time, trusting that they will work and that they will be driven safely. I still have my money in a bank, and I also hire an accountant and sign my tax forms without reading them. So I’m a hypocrite, big-time.
There’s another thing I should clear up: I’m not claiming I understand everything about climate research just because I talked to an expert for 2 or 3 hours. I am certainly not an expert, nor am I planning to become one. Even so, I did learn a lot, and the research I undertook was incredibly useful to me.
So, for example, my father is a climate change denier, and I have heard him give a list of scientific facts to argue against climate change. I asked my expert to counter-argue these points, and he did so. I also asked him to explain the underlying model at a high level, which he did.
My conclusion wasn’t that I’ve looked carefully into the model and it’s right, because that’s not possible in such a short time. My conclusion was that this guy is trustworthy and uses logical argument, which he’s happy to share with interested people, and moreover he manages to defend against deniers without being intellectually defensive. In the end, I’m trusting him, an expert.
On the other hand, if I met another person with a totally different conclusion, who also impressed me as intellectually honest and curious, then I’d definitely listen to that guy too, and I’d be willing to change my mind.
So I do imbue models and theories with a limited amount of trust depending on how much sense they makes to me. I think that’s reasonable, and it’s in line with my advocacy of scientific interpreters. Obviously not all scientific interpreters would be telling the same story, but that’s not important – in fact it’s vital that they don’t, because it is a privilege to be allowed to listen to the different sides and be engaged in the debate.
If I sat down with an expert for a whole day, like my friend Jordan suggests, to determine if they were “right” on an issue where there’s argument among experts, then I’d fail, but even understanding what they were arguing about would be worthwhile and educational.
Let me say this another way: experts argue about what they don’t agree on, of course, since it would be silly for them to talk about what they do agree on. But it’s their commonality that we, the laypeople, are missing. And that commonality is often so well understood that we could understand it rather quickly if it was willingly explained to us. That would be a huge step.
So I wasn’t lying after all, if I am allowed to define the “it” that I did get at in the two hours with an expert. When I say I understood it, I didn’t mean everything, I meant a much larger chunk of the approach and method than I’d had before, and enough to evoke (limited) trust.
Something I haven’t addressed, which I need to think about more (please help!), is the question of what subjects require active skepticism. On of my commenters, Paul Stevens, brought this up:
… For me, lay people means John Q Public – public opinion because public opinion can shape policy. In practice, this only matters for a select few issues, such as climate change or science education. There is no impact to a lay person not understanding / believing in the Higgs particle for example.
Stephanie asks three important questions about trusting experts, which I paraphrase here:
- What does it take to look into a model yourself? How deeply must you probe?
- How do you avoid being manipulated when you do so?
- Why should we bother since stuff is so hard and we each have a limited amount of time?
I must confess I find the first two questions really interesting and I want to think about them, but I have a very little patience with the last question.
- I’ve seen too many people (individual modelers) intentionally deflect investigations into models by setting them up as so hard that it’s not worth it (or at least it seems not worth it). They use buzz words and make it seem like there’s a magical layer of their model which makes it too difficult for mere mortals. But my experience (as an arrogant, provocative, and relentless questioner) is that I can always understand a given model if I’m talking to someone who really understands it and actually wants to communicate it.
- It smacks of an excuse rather than a reason. If it’s our responsibility to understand something, then by golly we should do it, even if it’s hard.
- Too many things are left up to people whose intentions are not reasonable using this “too hard” argument, and it gives those people reason to make entire systems seem too difficult to penetrate. For a great example, see the financial system, which is consistently too complicated for regulators to properly regulate.
I’m sure I seem unbelievably cynical here, but that’s where I got by working in finance, where I saw first-hand how manipulative and manipulated mathematical modeling can become. And there’s no reason at all such machinations wouldn’t translate to the world of big data or climate modeling.
Speaking of climate modeling: first, it annoys me that people are using my “distrust the experts” line to be cast doubt on climate modelers.
People: I’m not asking you to simply be skeptical, I’m saying you should look into the models yourself! It’s the difference between sitting on a couch and pointing at a football game on TV and complaining about a missed play and getting on the football field yourself and trying to figure out how to throw the ball. The first is entertainment but not valuable to anyone but yourself. You are only adding to the discussion if you invest actual thoughtful work into the matter.
To that end, I invited an expert climate researcher to my house and asked him to explain the climate models to me and my husband, and although I’m not particularly skeptical of climate change research (more on that below when I compare incentives of the two sides), I asked obnoxious, relentless questions about the model until I was satisfied. And now I am satisfied. I am considering writing it up as a post.
As an aside, if climate researchers are annoyed by the skepticism, I can understand that, since football fans are an obnoxious group, but they should not get annoyed by people who want to actually do the work to understand the underlying models.
Another thing about climate research. People keep talking about incentives, and yes I agree wholeheartedly that we should follow the incentives to understand where manipulation might be taking place. But when I followed the incentives with respect to climate modeling, they bring me straight to climate change deniers, not to researchers.
Do we really think these scientists working with their research grants have more at stake than multi-billion dollar international companies who are trying to ignore the effect of their polluting factories on the environment? People, please. The bulk of the incentives are definitely with the business owners. Which is not to say there are no incentives on the other side, since everyone always wants to feel like their research is meaningful, but let’s get real.
I like this idea Stephanie comes up with:
Some sociologists of science suggest that translational “experts”–that is, “experts” who aren’t necessarily producing new information and research, but instead are “expert” enough to communicate stuff to those not trained in the area–can help bridge this divide without requiring everyone to become “experts” themselves. But that can also raise the question of whether these translational experts have hidden agendas in some way. Moreover, one can also raise questions of whether a partial understanding of the model might in some instances be more misleading than not looking into the model at all–examples of that could be the various challenges to evolution based on fairly minor examples that when fully contextualized seem minor but may pop out to someone who is doing a less systematic inquiry.
First, I attempt to make my blog something like a platform for this, and I also do my best to make my agenda not at all hidden so people don’t have to worry about that.
This raises a few issues for me:
- Right now we depend mostly on press to do our translations, but they aren’t typically trained as scientists. Does that make them more prone to being manipulated? I think it does.
- How do we encourage more translational expertise to emerge from actual experts? Currently, in academia, the translation to the general public of one’s research is not at all encouraged or rewarded, and outside academia even less so.
- Like Stephanie, I worry about hidden agendas and partial understandings, but I honestly think they are secondary to getting a robust system of translation started to begin with, which would hopefully in turn engage the general public with the scientific method and current scientific knowledge. In other words, the good outweighs the bad here.
I’ve really enjoyed the discussion on my post from yesterday about MOOCs and how I predict they are going to affect the education world. I could be wrong, of course, but I think this stuff is super interesting to think about.
One thing I thought about since writing the post yesterday, in terms of math departments, is that I used to urge people involved in math departments to be attentive to their calculus teaching.
The threat, as I saw it then, was this: if math departments are passive and boring and non-reactive about how they teach calculus, then other departments which need calculus for their majors would pick up the slack and we’d see calculus taught in economics, physics, and engineering departments.
The reason math departments should care about this is that calculus is the bread and butter of math departments – math departments in other countries who have lost calculus to other departments are very small. If you only need to teach math majors, it doesn’t require that many people to do that.
But now I don’t even bother saying this, because the threat from MOOCs is much bigger and is going to have a more profound effect, and moreover there’s nothing math departments can do to stop it. Well, they can bury their head in the sand but I don’t recommend it.
Once there’s a really good calculus sequence out there, why would departments continue to teach the old fashioned way? Once there’s a fantastic calculus-for-physics MOOC, or calculus-for-economics MOOC available, one would hope that math departments would admit they can’t do better.
Instead of the old-fashioned calculus approach they’d figure out a way to incorporate the MOOC and supplement it by forming study groups and leading sections on the material. This would require a totally different set-up, and probably fewer mathematicians.
Another thing. I think I’ve identified a few separate issues in the discussion that it makes sense to highlight. There are four things (at least) that are all rolled together in our current college and university experience:
- learning itself,
- research, and
So, MOOCs directly address learning but clearly want to control something about credentialing too, which I think won’t necessarily work. They also affect research because the role of professor as learning instructor will change. They give us nothing in terms of socializing.
But as commenters have pointed out, socializing students is a huge part of the college experience, and may be even more important than credentialing. Or another way of saying that is people look at your resume not so much to know what you know but to know how you’ve been socialized.
It makes me wonder how we will address the “socializing” part of education in the future. And it also makes me wonder where research will be in 100 years.
In the final week of Rachel Schutt’s Columbia Data Science course we heard from two groups of students as well as from Rachel herself.
Data Science; class consciousness
The first team of presenters consisted of Yegor, Eurry, and Adam. Many others whose names I didn’t write down contributed to the research, visualization, and writing.
First they showed us the very cool graphic explaining how self-reported skills vary by discipline. The data they used came from the class itself, which did this exercise on the first day:
so the star in the middle is the average for the whole class, and each star along the side corresponds to the average (self-reported) skills of people within a specific discipline. The dotted lines on the outside stars shows the “average” star, so it’s easier to see how things vary per discipline compared to the average.
Surprises: Business people seem to think they’re really great at everything except communication. Journalists are better at data wrangling than engineers.
We will get back to the accuracy of self-reported skills later.
We were asked, do you see your reflection in your star?
Also, take a look at the different stars. How would you use them to build a data science team? Would you want people who are good at different skills? Is it enough to have all the skills covered? Are there complementary skills? Are the skills additive, or do you need overlapping skills among team members?
If all data which had ever been collected were freely available to everyone, would we be better off?
Some ideas were offered:
- all nude photos are included. [Mathbabe interjects: it's possible to not let people take nude pics of you. Just sayin'.]
- so are passwords, credit scores, etc.
- how do we make secure transactions between a person and her bank considering this?
- what does it mean to be “freely available” anyway?
The data of power; the power of data
You see a lot of people posting crap like this on Facebook:
But here’s the thing: the Berner Convention doesn’t exist. People are posting this to their walls because they care about their privacy. People think they can exercise control over their data but they can’t. Stuff like this give one a false sense of security.
In Europe the privacy laws are stricter, and you can request data from Irish Facebook and they’re supposed to do it, but it’s still not easy to successfully do.
And it’s not just data that’s being collected about you – it’s data you’re collecting. As scientists we have to be careful about what we create, and take responsibility for our creations.
As Francois Rabelais said,
Wisdom entereth not into a malicious mind, and science without conscience is but the ruin of the soul.
Or as Emily Bell from Columbia said,
Every algorithm is editorial.
We can’t be evil during the day and take it back at hackathons at night. Just as journalists need to be aware that the way they report stories has consequences, so do data scientists. As a data scientist one has impact on people’s lives and how they think.
Here are some takeaways from the course:
- We’ve gained significant powers in this course.
- In the future we may have the opportunity to do more.
- With data power comes data responsibility.
Who does data science empower?
The second presentation was given by Jed and Mike. Again, they had a bunch of people on their team helping out.
Let’s start with a quote:
“Anything which uses science as part of its name isn’t political science, creation science, computer science.”
- Hal Abelson, MIT CS prof
Keeping this in mind, if you could re-label data science, would you? What would you call it?
Some comments from the audience:
- Let’s call it “modellurgy,” the craft of beating mathematical models into shape instead of metal
- Let’s call it “statistics”
Does it really matter what data science is? What should it end up being?
Chris Wiggins from Columbia contends there are two main views of what data science should end up being. The first stems from John Tukey, inventor of the fast fourier transform and the box plot, and father of exploratory data analysis. Tukey advocated for a style of research he called “data analysis”, emphasizing the primacy of data and therefore computation, which he saw as part of statistics. His descriptions of data analysis, which he saw as part of doing statistics, are very similar to what people call data science today.
The other prespective comes from Jim Gray, Computer Scientist from Microsoft. He saw the scientific ideals of the enlightenment age as expanding and evolving. We’ve gone from the theories of Darwin and Newton to experimental and computational approaches of Turing. Now we have a new science, a data-driven paradigm. It’s actually the fourth paradigm of all the sciences, the first three being experimental, theoretical, and computational. See more about this here.
Wait, can data science be both?
Note it’s difficult to stick Computer Science and Data Science on this line.
Statistics is a tool that everyone uses. Data science also could be seen that way, as a tool rather than a science.
Who does data science?
Here’s a graphic showing the make-up of Kaggle competitors. Teams of students collaborated to collect, wrangle, analyze and visualize this data:
The size of the blocks correspond to how many people in active competitions have an education background in a given field. We see that almost a quarter of competitors are computer scientists. The shading corresponds to how often they compete. So we see the business finance people do more competitions on average than the computer science people.
Consider this: the only people doing math competitions are math people. If you think about it, it’s kind of amazing how many different backgrounds are represented above.
We got some cool graphics created by the students who collaborated to get the data, process it, visualize it and so on.
Which universities offer courses on Data Science?
There will be 26 universities in total by 2013 that offer data science courses. The balls are centered at the center of gravity of a given state, and the balls are bigger if there are more in that state.
Where are data science jobs available?
- We see more professional schools offering data science courses on the west coast.
- It would also would be interesting to see this corrected for population size.
- Only two states had no jobs.
- Massachusetts #1 per capita, then Maryland
McKinsey says there will be hundreds of thousands of data science jobs in the next few years. There’s a massive demand in any case. Some of us will be part of that. It’s up to us to make sure what we’re doing is really data science, rather than validating previously held beliefs.
We need to advance human knowledge if we want to take the word “scientist” seriously.
How did this class empower you?
You are one of the first people to take a data science class. There’s something powerful there.
Thank you Rachel!
Last Day of Columbia Data Science Class, What just happened? from Rachel’s perspective
Recall the stated goals of this class were:
- learn about what it’s like to be a data scientists
- be able to do some of what a data scientist does
Hey we did this! Think of all the guest lectures; they taught you a lot of what it’s like to be a data scientist, which was goal 1. Here’s what I wanted you guys to learn before the class started based on what a data scientist does, and you’ve learned a lot of that, which was goal 2:
Mission accomplished! Mission accomplished?
Thought experiment that I gave to myself last Spring
How would you design a data science class?
Comments I made to myself:
- It’s not a well-defined body of knowledge, subject, no textbook!
- It’s popularized and celebrated in the press and media, but there’s no “authority” to push back
- I’m intellectually disturbed by idea of teaching a course when the body of knowledge is ill-defined
- I didn’t know who would show up, and what their backgrounds and motivations would be
- Could it become redundant with a machine learning class?
I asked questions of myself and from other people. I gathered information, and endured existential angst about data science not being a “real thing.” I needed to give it structure.
Then I started to think about it this way: while I recognize that data science has the potential to be a deep research area, it’s not there yet, and in order to actually design a class, let’s take a pragmatic approach: Recognize that data science exists. After all, there are jobs out there. I want to help students to be qualified for them. So let me teach them what it takes to get those jobs. That’s how I decided to approach it.
In other words, from this perspective, data science is what data scientists do. So it’s back to the list of what data scientists do. I needed to find structure on top of that, so the structure I used as a starting point were the data scientist profiles.
Data scientist profiles
This was a way to think about your strengths and weaknesses, as well as a link between speakers. Note it’s easy to focus on “technical skills,” but it can also be problematic in being too skills-based, as well as being problematic because it has no scale, and no notion of expertise. On the other hand it’s good in that it allows for and captures variability among data scientists.
I assigned weekly guest speakers topics related to their strengths. We held lectures, labs, and (optional) problem sessions. From this you got mad skillz:
- programming in R
- some python
- you learned some best practices about coding
From the perspective of machine learning,
- you know a bunch of algorithms like linear regression, logistic regression, k-nearest neighbors, k-mean, naive Bayes, random forests,
- you know what they are, what they’re used for, and how to implement them
- you learned machine learning concepts like training sets, test sets, over-fitting, bias-variance tradeoff, evaluation metrics, feature selection, supervised vs. unsupervised learning
- you learned about recommendation systems
- you’ve entered a Kaggle competition
Importantly, you now know that if there is an algorithm and model that you don’t know, you can (and will) look it up and figure it out. I’m pretty sure you’ve all improved relative to how you started.
You’ve learned some data viz by taking flowing data tutorials.
You’ve learned statistical inference, because we discussed
- observational studies,
- causal inference, and
- experimental design.
- We also learned some maximum likelihood topics, but I’d urge you to take more stats classes.
In the realm of data engineering,
- we showed you map reduce and hadoop
- we worked with 30 separate shards
- we used an api to get data
- we spent time cleaning data
- we’ve processed different kinds of data
As for communication,
- you wrote thoughts in response to blog posts
- you observed how different data scientists communicate or present themselves, and have different styles
- your final project required communicating among each other
As for domain knowledge,
- lots of examples were shown to you: social networks, advertising, finance, pharma, recommender systems, dallas art museum
I heard people have been asking the following: why didn’t we see more data science coming from non-profits, governments, and universities? Note that data science, the term, was born in for-profits. But the truth is I’d also like to see more of that. It’s up to you guys to go get that done!
How do I measure the impact of this class I’ve created? Is it possible to incubate awesome data science teams in the classroom? I might have taken you from point A to point B but you might have gone there anyway without me. There’s no counterfactual!
Can we set this up as a data science problem? Can we use a causal modeling approach? This would require finding students who were more or less like you but didn’t take this class and use propensity score matching. It’s not a very well-defined experiment.
But the goal is important: in industry they say you can’t learn data science in a university, that it has to be on the job. But maybe that’s wrong, and maybe this class has proved that.
What has been the impact on you or to the outside world? I feel we have been contributing to the broader discourse.
Does it matter if there was impact? and does it matter if it can be measured or not? Let me switch gears.
What is data science again?
Data science could be defined as:
- A set of best practices used in tech companies, which is how I chose to design the course
- A space of problems that could be solved with data
- A science of data where you can think of the data itself as units
The bottom two have the potential to be the basis of a rich and deep research discipline, but in many cases, the way the term is currently used is:
- Pure hype
But it doesn’t matter how we define it, as much as that I want for you:
- to be problem solvers
- to be question askers
- to think about your process
- to use data responsibly and make the world better, not worse.
More on being problem solvers: cultivate certain habits of mind
Here’s a possible list of things to strive for, taken from here:
Here’s the thing. Tons of people can implement k-nearest neighbors, and many do it badly. What matters is that you cultivate the above habits, remain open to continuous learning.
In education in traditional settings, we focus on answers. But what we probably should focus on is how a student behaves when they don’t know the answer. We need to have qualities that help us find the answer.
How would you design a data science class around habits of mind rather than technical skills? How would you quantify it? How would you evaluate? What would students be able to write on their resumes?
Comments from the students:
- You’d need to keep making people doing stuff they don’t know how to do while keeping them excited about it.
- have people do stuff in their own domains so we keep up wonderment and awe.
- You’d use case studies across industries to see how things work in different contexts
More on being question-askers
Some suggestions on asking questions of others:
- start with assumption that you’re smart
- don’t assume the person you’re talking to knows more or less. You’re not trying to prove anything.
- be curious like a child, not worried about appearing stupid
- ask for clarification around notation or terminology
- ask for clarification around process: where did this data come from? how will it be used? why is this the right data to use? who is going to do what? how will we work together?
Some questions to ask yourself
- does it have to be this way?
- what is the problem?
- how can I measure this?
- what is the appropriate algorithm?
- how will I evaluate this?
- do I have the skills to do this?
- how can I learn to do this?
- who can I work with? Who can I ask?
- how will it impact the real world?
Data Science Processes
In addition to being problem-solvers and question-askers, I mentioned that I want you to think about process. Here are a couple processes we discussed in this course:
(1) Real World –> Generates Data –>
–> Collect Data –> Clean, Munge (90% of your time)
–> Exploratory Data Analysis –>
–> Feature Selection –>
–> Build Model, Build Algorithm, Visualize
–> Evaluate –>Iterate–>
–> Impact Real World
(2) Asking questions of yourselves and others –>
Identifying problems that need to be solved –>
Gathering information, Measuring –>
Learning to find structure in unstructured situations–>
Framing Problem –>
Creating Solutions –> Evaluating
Come up with a business that improves the world and makes money and uses data
Comments from the students:
- autonomous self-driving cars you order with a smart phone
- find all the info on people and then show them how to make it private
- social network with no logs and no data retention
10 Important Data Science Ideas
Of all the blog posts I wrote this semester, here’s one I think is important:
Confidence and Uncertainty
Let’s talk about confidence and uncertainty from a couple perspectives.
First, remember that statistical inference is extracting information from data, estimating, modeling, explaining but also quantifying uncertainty. Data Scientists could benefit from understanding this more. Learn more statistics and read Ben’s blog post on the subject.
Second, we have the Dunning-Kruger Effect.
Have you ever wondered why don’t people say “I don’t know” when they don’t know something? This is partly explained through an unconscious bias called the Dunning-Kruger effect.
Basically, people who are bad at something have no idea that they are bad at it and overestimate their confidence. People who are super good at something underestimate their mastery of it. Actual competence may weaken self-confidence.
Design an app to combat the dunning-kruger effect.
Optimizing your life, Career Advice
What are you optimizing for? What do you value?
- money, need some minimum to live at the standard of living you want to, might even want a lot.
- time with loved ones and friends
- doing good in the world
- personal fulfillment, intellectual fulfillment
- goals you want to reach or achieve
- being famous, respected, acknowledged
- some weighted function of all of the above. what are the weights?
What constraints are you under?
- external factors (factors outside of your control)
- your resources: money, time, obligations
- who you are, your education, strengths & weaknesses
- things you can or cannot change about yourself
There are many possible solutions that optimize what you value and take into account the constraints you’re under.
So what should you do with your life?
Remember that whatever you decide to do is not permanent so don’t feel too anxious about it, you can always do something else later –people change jobs all the time
But on the other hand, life is short, so always try to be moving in the right direction (optimizing for what you care about).
If you feel your way of thinking or perspective is somehow different than what those around you are thinking, then embrace and explore that, you might be onto something.
I’m always happy to talk to you about your individual case.
Next Gen Data Scientists
The second blog post I think is important is this “manifesto” that I wrote:
Next-Gen Data Scientists. That’s you! Go out and do awesome things, use data to solve problems, have integrity and humility.
Here’s our class photo!
This week’s guest lecturer in Rachel Schutt’s Columbia Data Science class was Claudia Perlich. Claudia has been the Chief Scientist at m6d for 3 years. Before that she was a data analytics group at the IBM center that developed Watson, the computer that won Jeopardy!, although she didn’t work on that project. Claudia got her Ph.D. in information systems at NYU and now teaches a class to business students in data science, although mostly she addresses how to assess data science work and how to manage data scientists. Claudia also holds a masters in Computer Science.
Claudia is a famously successful data mining competition winner. She won the KDD Cup in 2003, 2007, 2008, and 2009, the ILP Challenge in 2005, the INFORMS Challenge in 2008, and the Kaggle HIV competition in 2010.
She’s also been a data mining competition organizer, first for the INFORMS Challenge in 2009 and then for the Heritage Health Prize in 2011. Claudia claims to be retired from competition.
Claudia’s advice to young people: pick your advisor first, then choose the topic. It’s important to have great chemistry with your advisor, and don’t underestimate the importance.
Here’s what Claudia historically does with her time:
- predictive modeling
- data mining competitions
- publications in conferences like KDD and journals
- digging around data (her favorite part)
Claudia likes to understand something about the world by looking directly at the data.
Here’s Claudia’s skill set:
- plenty of experience doing data stuff (15 years)
- data intuition (for which one needs to get to the bottom of the data generating process)
- dedication to the evaluation (one needs to cultivate a good sense of smell)
- model intuition (we use models to diagnose data)
Claudia also addressed being a woman. She says it works well in the data science field, where her intuition is useful and is used. She claims her nose is so well developed by now that she can smell it when something is wrong. This is not the same thing as being able to prove something algorithmically. Also, people typically remember her because she’s a woman, even when she don’t remember them. It has worked in her favor, she says, and she’s happy to admit this. But then again, she is where she is because she’s good.
Someone in the class asked if papers submitted for journals and/or conferences are blind to gender. Claudia responded that it was, for some time, typically double-blind but now it’s more likely to be one-sided. And anyway there was a cool analysis that showed you can guess who wrote a paper with 80% accuracy just by knowing the citations. So making things blind doesn’t really help. More recently the names are included, and hopefully this doesn’t make things too biased. Claudia admits to being slightly biased towards institutions – certain institutions prepare better work.
Skills and daily life of a Chief Data Scientist
Claudia’s primary skills are as follows:
- Data manipulation: unix (sed, awk, etc), Perl, SQL
- Modeling: various methods (logistic regression, nearest neighbors, k-nearest neighbors, etc)
- Setting things up
She mentions that the methods don’t matter as much as how you’ve set it up, and how you’ve translated it into something where you can solve a question.
More recently, she’s been told that at work she spends:
- 40% of time as “contributor”: doing stuff directly with data
- 40% of time as “ambassador”: writing stuff, giving talks, mostly external communication to represent m6d, and
- 20% of time in “leadership” of her data group
At IBM it was much more focused in the first category. Even so, she has a flexible schedule at m6d and is treated well.
The goals of the audience
She asked the class, why are you here? Do you want to:
- become a data scientist? (good career choice!)
- work with data scientist?
- work for a data scientist?
- manage a data scientist?
Most people were trying their hands at the first, but we had a few in each category.
She mentioned that it matters because the way she’d talk to people wanting to become a data scientist would be different from the way she’d talk to someone who wants to manage them. Her NYU class is more like how to manage one.
So, for example, you need to be able to evaluate their work. It’s one thing to check a bubble sort algorithm or check whether a SQL server is working, but checking a model which purports to give the probability of people converting is different kettle of fish.
For example, try to answer this: how much better can that model get if you spend another week on it? Let’s face it, quality control is hard for yourself as a data miner, so it’s definitely hard for other people. There’s no easy answer.
There’s an old joke that comes to mind: What’s the difference between the scientist and a consultant? The scientists asks, how long does it take to get this right? whereas the consultant asks, how right can I get this in a week?
Insights into data
A student asks, how do you turn a data analysis into insights?
For example, decision trees you interpret, and people like them because they’re easy to interpret, but I’d ask, why does it look like it does? A slightly different data set would give you a different tree and you’d get a different conclusion. This is the illusion of understanding. I tend to be careful with delivering strong insights in that sense.
Data mining competitions
Claudia drew a distinction between different types of data mining competitions.
On the one hand you have the “sterile” kind, where you’re given a clean, prepared data matrix, a standard error measure, and where the features are often anonymized. This is a pure machine learning problem.
Examples of this first kind are: KDD Cup 2009 and 2011 (Netflix). In such competitions, your approach would emphasize algorithms and computation. The winner would probably have heavy machines and huge modeling ensembles.
On the other hand, you have the “real world” kind of data mining competition, where you’re handed raw data, which is often in lots of different tables and not easily joined, where you set up the model yourself and come up with task-specific evaluations. This kind of competition simulates real life more.
Examples of this second kind are: KDD cup 2007, 2008, and 2010. If you’re competing in this kind of competition your approach would involve understanding the domain, analyzing the data, and building the model. The winner might be the person who best understands how to tailor the model to the actual question.
Claudia prefers the second kind, because it’s closer to what you do in real life. In particular, the same things go right or go wrong.
How to be a good modeler
Claudia claims that data and domain understanding is the single most important skill you need as a data scientist. At the same time, this can’t really be taught – it can only be cultivated.
A few lessons learned about data mining competitions that Claudia thinks are overlooked in academics:
- Leakage: the contestants best friend and the organizers/practitioners worst nightmare. There’s always something wrong with the data, and Claudia has made an artform of figuring out how the people preparing the competition got lazy or sloppy with the data.
- Adapting learning to real-life performance measures beyond standard measures like MSE, error rate, or AUC (profit?)
- Feature construction/transformation: real data is rarely flat (i.e. given to you in a beautiful matrix) and good, practical solutions for this problem remains a challenge.
Leakage refers to something that helps you predict something that isn’t fair. It’s a huge problem in modeling, and not just for competitions. Oftentimes it’s an artifact of reversing cause and effect.
Example 1: There was a competition where you needed to predict S&P in terms of whether it would go up or go down. The winning entry had a AUC (area under the ROC curve) of 0.999 out of 1. Since stock markets are pretty close to random, either someone’s very rich or there’s something wrong. There’s something wrong.
In the good old days you could win competitions this way, by finding the leakage.
Example 2: Amazon case study: big spenders. The target of this competition was to predict customers who spend a lot of money among customers using past purchases. The data consisted of transaction data in different categories. But a winning model identified that “Free Shipping = True” was an excellent predictor
What happened here? The point is that free shipping is an effect of big spending. But it’s not a good way to model big spending, because in particular it doesn’t work for new customers or for the future. Note: timestamps are weak here. The data that included “Free Shipping = True” was simultaneous with the sale, which is a no-no. We need to only use data from beforehand to predict the future.
Example 3: Again an online retailer, this time the target is predicting customers who buy jewelry. The data consists of transactions for different categories. A very successful model simply noted that if sum(revenue) = 0, then it predicts jewelry customers very well?
What happened here? The people preparing this data removed jewelry purchases, but only included people who bought something in the first place. So people who had sum(revenue) = 0 were people who only bought jewelry. The fact that you only got into the dataset if you bought something is weird: in particular, you wouldn’t be able to use this on customers before they finished their purchase. So the model wasn’t being trained on the right data to make the model useful. This is a sampling problem, and it’s common.
Example 4: This happened at IBM. The target was to predict companies who would be willing to buy “websphere” solutions. The data was transaction data + crawled potential company websites. The winning model showed that if the term “websphere” appeared on the company’s website, then they were great candidates for the product.
What happened? You can’t crawl the historical web, just today’s web.
You’re trying to study who has breast cancer. The patient ID, which seemed innocent, actually has predictive power. What happened?
In the above image, red means cancerous, green means not. it’s plotted by patient ID. We see three or four distinct buckets of patient identifiers. It’s very predictive depending on the bucket. This is probably a consequence of using multiple databases, some of which correspond to sicker patients are more likely to be sick.
A student suggests: for the purposes of the contest they should have renumbered the patients and randomized.
Claudia: would that solve the problem? There could be other things in common as well.
A student remarks: The important issue could be to see the extent to which we can figure out which dataset a given patient came from based on things besides their ID.
Claudia: Think about this: what do we want these models for in the first place? How well can you predict cancer?
Given a new patient, what would you do? If the new patient is in a fifth bin in terms of patient ID, then obviously don’t use the identifier model. But if it’s still in this scheme, then maybe that really is the best approach.
This discussion brings us back to the fundamental problem that we need to know what the purpose of the model is and how is it going to be used in order to decide how to do it and whether it’s working.
During an INFORMS competition on pneumonia predictions in hospital records, where the goal was to predict whether a patient has pneumonia, a logistic regression which included the number of diagnosis codes as a numeric feature (AUC of 0.80) didn’t do as well as the one which included it as a categorical feature (0.90). What’s going on?
This had to do with how the person prepared the data for the competition:
The diagnosis code for pneumonia was 486. So the preparer removed that (and replaced it by a “-1″) if it showed up in the record (rows are different patients, columns are different diagnoses, there are max 4 diagnoses, “-1″ means there’s nothing for that entry).
Moreover, to avoid telling holes in the data, the preparer moved the other diagnoses to the left if necessary, so that only “-1″‘s were on the right.
There are two problems with this:
- If the column has only “-1″‘s, then you know it started out with only pneumonia, and
- If the column has no “-1″‘s, you know there’s no pneumonia (unless there are actually 5 diagnoses, but that’s less common).
This was enough information to win the competition.
Note: winning competition on leakage is easier than building good models. But even if you don’t explicitly understand and game the leakage, your model will do it for you. Either way, leakage is a huge problem.
How to avoid leakage
Claudia’s advice to avoid this kind of problem:
- You need a strict temporal cutoff: remove all information just prior to the event of interest (patient admission).
- There has to be a timestamp on every entry and you need to keep
- Removing columns asks for trouble
- Removing rows can introduce inconsistencies with other tables, also causing trouble
- The best practice is to start from scratch with clean, raw data after careful consideration
- You need to know how the data was created! I only work with data I pulled and prepared myself (or maybe Ori).
How do I know that my model is any good?
With powerful algorithms searching for patterns of models, there is a serious danger of over fitting. It’s a difficult concept, but the general idea is that “if you look hard enough you’ll find something” even if it does not generalize beyond the particular training data.
To avoid overfitting, we cross-validate and we cut down on the complexity of the model to begin with. Here’s a standard picture (although keep in mind we generally work in high dimensional space and don’t have a pretty picture to look at):
The picture on the left is underfit, in the middle is good, and on the right is overfit.
The model you use matters when it concerns overfitting:
So for the above example, unpruned decision trees are the most over fitting ones. This is a well-known problem with unpruned decision trees, which is why people use pruned decision trees.
Claudia dismisses accuracy as a bad evaluation method. What’s wrong with accuracy? It’s inappropriate for regression obviously, but even for classification, if the vast majority is of binary outcomes are 1, then a stupid model can be accurate but not good (guess it’s always “1″), and a better model might have lower accuracy.
Probabilities matter, not 0′s and 1′s.
Nobody makes decisions on binary outcomes. I want to know the probability I have breast cancer, I don’t want to be told yes or no. It’s much more information. I care about probabilities.
How to evaluate a probability model
We separately evaluate the ranking and the calibration. To evaluate the ranking, we use the ROC curve and calculate the area under it, typically ranges from 0.5-1.0. This is independent of scaling and calibration. Here’s an example of how to draw an ROC curve:
Sometimes to measure rankings, people draw the so-called lift curve:
The key here is that the lift is calculated with respect to a baseline. You draw it at a given point, say 10%, by imagining that 10% of people are shown ads, and seeing how many people click versus if you randomly showed 10% of people ads. A lift of 3 means it’s 3 times better.
How do you measure calibration? Are the probabilities accurate? If the model says probability of 0.57 that I have cancer, how do I know if it’s really 0.57? We can’t measure this directly. We can only bucket those predictions and then aggregately compare those in that prediction bucket (say 0.50-0.55) to the actual results for that bucket.
For example, here’s what you get when your model is an unpruned decision tree, where the blue diamonds are buckets:
A good model would show buckets right along the x=y curve, but here we’re seeing that the predictions were much more extreme than the actual probabilities. Why does this pattern happen for decision trees?
Claudia says that this is because trees optimize purity: it seeks out pockets that have only positives or negatives. Therefore its predictions are more extreme than reality. This is generally true about decision trees: they do not generally perform well with respect to calibration.
Logistic regression looks better when you test calibration, which is typical:
- Accuracy is almost never the right evaluation metric.
- Probabilities, not binary outcomes.
- Separate ranking from calibration.
- Ranking you can measure with nice pictures: ROC, lift
- Calibration is measured indirectly through binning.
- Different models are better than others when it comes to calibration.
- Calibration is sensitive to outliers.
- Measure what you want to be good at.
- Have a good baseline.
Choosing an algorithm
This is not a trivial question and in particular small tests may steer you wrong, because as you increase the sample size the best algorithm might vary: often decision trees perform very well but only if there’s enough data.
In general you need to choose your algorithm depending on the size and nature of your dataset and you need to choose your evaluation method based partly on your data and partly on what you wish to be good at. Sum of squared error is maximum likelihood loss function if your data can be assumed to be normal, but if you want to estimate the median, then use absolute errors. If you want to estimate a quantile, then minimize the weighted absolute error.
We worked on predicting the number of ratings of a movie will get in the next year, and we assumed a poisson distributions. In this case our evaluation method doesn’t involve minimizing the sum of squared errors, but rather something else which we found in the literature specific to the Poisson distribution, which depends on the single parameter :
Charity direct mail campaign
Let’s put some of this together.
Say we want to raise money for a charity. If we send a letter to every person in the mailing list we raise about $9000. We’d like to save money and only send money to people who are likely to give – only about 5% of people generally give. How can we do that?
If we use a (somewhat pruned, as is standard) decision tree, we get $0 profit: it never finds a leaf with majority positives.
If we use a neural network we still make only $7500, even if we only send a letter in the case where we expect the return to be higher than the cost.
This looks unworkable. But if you model is better, it’s not. A person makes two decisions here. First, they decide whether or not to give, then they decide how much to give. Let’s model those two decisions separately, using:
Note we need the first model to be well-calibrated because we really care about the number, not just the ranking. So we will try logistic regression for first half. For the second part, we train with special examples where there are donations.
Altogether this decomposed model makes a profit of $15,000. The decomposition made it easier for the model to pick up the signals. Note that with infinite data, all would have been good, and we wouldn’t have needed to decompose. But you work with what you got.
Moreover, you are multiplying errors above, which could be a problem if you have a reason to believe that those errors are correlated.
We are not meant to understand data. Data are outside of our sensory systems and there are very few people who have a near-sensory connection to numbers. We are instead meant to understand language.
We are not mean to understand uncertainty: we have all kinds of biases that prevent this from happening and are well-documented.
Modeling people in the future is intrinsically harder than figuring out how to label things that have already happened.
Even so we do our best, and this is through careful data generation, careful consideration of what our problem is, making sure we model it with data close to how it will be used, making sure we are optimizing to what we actually desire, and doing our homework in learning which algorithms fit which tasks.
I’m very happy to say I just signed a book contract with my co-author, Rachel Schutt, to publish a book with O’Reilly called Doing Data Science.
For those of you who’ve been reading along for free as I’ve been blogging it, there might not be a huge incentive to buy it, but I can promise you more and better math, more explicit usable formulas, some sample code, and an overall better and more thought-out narrative.
It’s supposed to be published in May with a possible early release coming up at the end of February, in time for the O’Reilly Strata Santa Clara conference, where Rachel will be speaking about it and about other stuff curriculum related. Hopefully people will pick it up in time to teach their data science courses in Fall 2013.
Speaking of Rachel, she’s also been selected to give a TedXWomen talk at Barnard on December 1st, which is super exciting. She’s talking about advocating for the social good using data. Unfortunately the event is invitation-only, otherwise I’d encourage you all to go and hear her words of wisdom. Update: word on the street is that it will be video-taped.