| Paper: |
MLSP-L1.5 |
| Session: |
Blind Signal Separation and Matrix Factorizations |
| Location: |
Room B-2 |
| Session Time: |
Thursday, March 29, 10:30 - 12:30 |
| Presentation Time: |
Thursday, March 29, 11:50 - 12:10 |
| Presentation: |
Lecture |
| Topic: |
|
| Paper Title: |
Dynamic Matrix Factorization: A State Space Approach |
| Authors: |
John Sun, Massachusetts Institute of Technology, United States; Kush Varshney, Karthik Subbian, IBM T.J. Watson Research Center, United States |
| Abstract: |
Matrix factorization from a small number of observed entries has recently garnered much attention as the key ingredient of successful recommendation systems. One unresolved problem in this area is how to adapt current methods to handle changing user preferences over time. Recent proposals to address this issue are heuristic in nature and do not fully exploit the time-dependent structure of the problem. As a principled and general temporal formulation, we propose a dynamical state space model of matrix factorization. Our proposal builds upon probabilistic matrix factorization, a Bayesian model with Gaussian priors. We utilize results in state tracking, i.e. the Kalman filter, to provide accurate recommendations in the presence of both process and measurement noise. We show how system parameters can be learned via expectation-maximization and provide comparisons to current published techniques. |