Bayesian Probabilistic Tensor Factorization

Intro

This pages gives you the Bayesian Probabilistic Tensor Factorization (BPTF) algorithm described in the following paper:

Liang Xiong, Xi Chen, Tzu-kuo Huang, Jeff Schneider, and Jaime Carbonell, Temporal Collaborative Filtering with Bayesian Probabilistic Tensor Factorization, SIAM Data Mining 2010 (SDM 10). [pdf]

As a gift/for comparison, the Probabilistic Matrix Factorization and Bayesian Probabilistic Matrix Factorization are also provided.

Code

Matlab Code

This demo of BPTF is written in Matlab with mex. To run it, you need Matlab, some CPP compiler that Matlab supports, and a fast machine. All you have to do is to downloaded the code, uncompress it, and run "demo.m".

In case you are interested in fast and parallel processing: the GraphLab contains an awesomely fast implementation of this algorithm. Check it out.

Misc

Here are some details you might be interested in:

• The data provided is extracted from 4% of the Netflix data set. 20% users and 20% movies were randomly selected from the whole pool. Each user or movie is just a coordinate in the tensor, and no other information is conveyed (If you find this still violates people's rights, let me know.). The time index in the tensor is based on calendar months, except that some early months were combined together.
• We represent a sparse tensor by a Matlab struct with fields: subs, vals, size. subs contains the coordinates of the entries, one row per entry. vals contains the values for these entries. And size is the dimension of the tensor.
• The intialization method used here is based on alternating least squares. Its prediction is not as good as the gradient-based one used in the paper, but it needs less tuning and runs faster.

Liang Xiong, 2010-11-15