Recommender systems & ranking
Matrix factorization models have been widely used in recommender systems. We are interested in developing novel algorithms for matrix factorization and scaling to massive datasets. Recently we are also interested in applying deep learning in recommender systems, and it's interpretability.
- Classical matrix factorization considers point-wise loss (e.g., measure the square error between predicted score and real score). However, point-wise loss may not reflect the real preference: for example, it's possible that Alice rates all the movies with score 3, 4, 5 while Bob rates all the movies withe score 1, 2, 3. We try to develop better loss functions for matrix factorization.
We develop a scalable solver on pair-wise ranking loss, and show it outperforms (and is more stable) than traditional matrix factorization, with similar training time.
We further apply the "list-wise loss" proposed in the learning-to-rank literature to personalized recommender system. We are able to make it as fast as point-wise and pair-wise models, and it performs much better in the implicit feedback setting.
- Online game analysis:
As a 14-year online MOBA game player (starting with Warcraft III dota & Sanguo), I'm also interested in analyzing game data. In our recent paper, we are interested in the ranking problem based on match results, but different from classical ranking, each time we only observe a group comparison instead of pair comparison result.
- We are also interested in incorporating side information (explicit features) in the latent feature models. See some of our earlier work:
Thanks to my students (Liwei Wu, Yao Li) and collaborators (James Sharpnack, Kai-Yang Chiang, Inderjit Dhillon, Kevin Fujii, Fushing Hsieh) for all of these interesting work.