Description |
Recommendation systems are commonly used in e-commerce to suggest relevant content to users. Popular examples include Amazon, iTunes Genius, Google News and typically they operate on a corpus of millions of users/items. Such recommendation systems often rely on "collaborative filters" - algorithms that use implicit or explicit user-item ratings to make recommendations. In this talk I will
a) discuss the current state-of-the-art in collaborative filtering;
b) show connections with the problem of channel coding arising in communications and exploit this viewpoint to derive performance limits for collaborative filters;
c) introduce a simple scalable algorithm based on the principle of "popularity amongst friends (PAF)";
d) show that PAF yields competitive performance on real life datasets such Movielens and Netflix movie ratings;
e) show that PAF is near-optimal in a certain regime.
|