Since I work with recommender systems, I’d hardly be doing my job if I didn’t notice things like Google Reader’s new feed recommendations. From the description of how the recommender works on the Google help page (which is unfortunately not very specific):

Your recommendations list is automatically generated. It takes into account the feeds you’re already subscribed to, as well as information from your Web History, including your location. Aggregated across many users, this information can indicate which feeds are popular among people with similar interests. For instance, if a lot of people subscribe to feeds about both peanut butter and jelly, and you only subscribe to feeds about peanut butter, Reader will recommend that you try some jelly.

This sounds like they are using a hybrid recommender system. When you are recommending items (in this case feeds) to users, you can either consider the qualities of the items themselves (content-based) or the behavior of people similar to you (collaborative filtering). The Netflix Prize is a collaborative filtering case for the most part, though it is possible to add in some amount of content.

So content-based recommenders typically do a good job of finding items that you will like that are similar to items you have already tried. It won’t necessarily do a good job with items that you will like but that aren’t like things you have already tried. Collaborative filtering does a good job of solving that problem, but then fails when you haven’t provided very much information about what you like. So usually when the two approaches are combined, the hybrid system that results is better.

Google’s system sounds like it is using a content-based strategy with the feeds that you are already subscribed to as well as your web history. With that, it probably builds a profile of what sort of user you are. It then compares you to the tens of thousands of other users who use Google Reader (surely < 1M users?) and finds the ones closest to you. This is your neighborhood. Here, Google offers no help determining what sort of machine learning technique is used to actually predict your preferences. Maybe none at all. Usually kNN (k-Nearest Neighbors) is used and the similarity measure is either cosine similarity or Pearson correlation, since we aren’t dealing with explicit ratings, but a vector of content features.

This is mostly wild speculation, but it’s fun for me, so hey. The ex-intern who worked on it is a math PhD student at MIT, Olga Stroilova. Unfortunately, I can’t find a web page for her that would hint at what sorts of recommender strategies she’s used in the past. Being math though, I’m thinking maybe something more interesting like SVD (Singular Value Decomposition) or MMMF (Maximum Margin Matrix Factorization). Then again, it could’ve just been straight kNN. It probably would do well enough.

I tried it out and it actually did a great job of suggesting some cool feeds. Many of them were things that were clearly similar to stuff I like but not really up my alley. I subscribe to some New Scientist feeds for technology and earth science, but not to the health and human body feeds. It suggested those.

The cool ones that I added were:

If you use Google Reader, let me know what sort of experience you have with the recommendations. I’ve been hearing about this all day from a number of sources [here, here, and here, to name a few].