MAA Distinguished Lecture Series: Start Your Own Netflix
I’m on my way to D.C. today to give an alleged “distinguished lecture” to a group of mathematics enthusiasts. I misspoke in a previous post where I characterized the audience to consist of math teachers. In fact, I’ve been told it will consist primarily of people with some mathematical background, with typically a handful of high school teachers, a few interested members of the public, and a number of high school and college students included in the group.
So I’m going to try my best to explain three different ways of approaching recommendation engine building for services such as Netflix. I’ll be giving high-level descriptions of a latent factor model (this movie is violent and we’ve noticed you like violent movies), of the co-visitation model (lots of people who’ve seen stuff you’ve seen also saw this movie) and the latent topic model (we’ve noticed you like movies about the Hungarian 1956 Revolution). Then I’m going to give some indication of the issues in doing these massive-scale calculation and how it can be worked out.
And yes, I double-checked with those guys over at Netflix, I am allowed to use their name as long as I make sure people know there’s no affiliation.
In addition to the actual lecture, the MAA is having me give a 10-minute TED-like talk for their website as well as an interview. I am psyched by how easy it is to prepare my slides for that short version using prezi, since I just removed a bunch of nodes on the path of the material without removing the material itself. I will make that short version available when it comes online, and I also plan to share the longer prezi publicly.
[As an aside, and not to sound like an advertiser for prezi (no affiliation with them either!), but they have a free version and the resulting slides are pretty cool. If you want to be able to keep your prezis private you have to pay, but not as much as you’d need to pay for powerpoint. Of course there’s always Open Office.]
Train reading: Wrong Answer: the case against Algebra II, by Nicholson Baker, which was handed to me emphatically by my friend Nick. Apparently I need to read this and have an opinion.
mathbabe 
Sounds like a great talk–looking forward to the video! Any chance you’d give this talk to a group of smart public high school students? 🙂
LikeLike
Collaborative Filtering is a remarkable bit of science with Filter Bubble (Eli Pariser) implications. I really like the whole concept of getting recommendations from a quasi empirical ‘system,’ but my experience has shown that it’s less effective for more eclectic personalities. The fact that it may route people down the same path over and over is a downside.
What gets really strange is when psycholgists create questionnaires, the meaning of which is largely inscrutable to the participant. Of course they profess to know even better what you mean for not having your intellect in the way. Psychological ‘factors’ in recommendation engines would seem similarly fraught with presumption and ambiguity.
You left out brute force user ratings – people who liked or disliked the same films to the same degree as you liked this one you haven’t seen. That approach wouldn’t care about the ‘why.’ You also left about the broader concept of ‘genres.’ Netflix (and other movie rating systems) seem to see many people’s tastes as genre-centric.
LikeLike
I’m giving this talk in December! What did they tell you about the audience? I was going to give a talk I’ve given before as a “public lecture.” High school and college students would be perfect for the level of lecture I’m planning to give, is that who it’ll mostly be?
LikeLike
No idea but I will report back.
LikeLike