It looks like it's just returning my watched repos -- what type of recommendation algorithm is it running?

It might service some of your watched repos because of data quality issues. It uses a very basic nearest neighbor shared item set model. I'm definitely planning on improving the quality of the results in the near future.

Try loading the data into Neo4j and running Gremlin on it (https://github.com/tinkerpop/gremlin/wiki).

Examples...

Ranked similarity search https://groups.google.com/d/msg/neo4j/tuoqDlq45ys/1g1G9nTzjh...

Multi-Item Based Recommendation https://groups.google.com/d/msg/neo4j/XlGGdsSedVw/7pCPqghA84...

Recommendation engine https://groups.google.com/d/msg/neo4j/_dnyanUSmtk/eDhJ6rjNCn...

There were a few interesting repos in there I hadn't heard of - nice work!

Coincidentally, I'm building a github recommendation service as a SNA research project. You may be aware, but for those who are not, GitHub actually had a NetFlix-style recommendation contest in 2009.

https://github.com/blog/466-the-2009-github-contest

That's awesome. I didn't know about that.

Neat in concept, but mostly useless in that it recommended pretty much all projects that I have zero interest in (javascript/coffeescript stuff), which is odd considering I'm not watching any Javascript projects, or at least I don't think I am. It made a couple of shell recommendations, one for IE which I would never touch, one for Spark which I already have installed, and one for Gitflow. Gitflow seems neat though, so I guess one good recommendation might have been worth the rest of the noise?

This is cool.