NN-CMF: Neural Network Coupled Matrix Factorization

Overview

How can we accurately predict item ratings by users using multiple coupled data? Nowadays huge amount of sparse information is available, and in many cases auxiliary data associated with rating data are also present. Training a model to predict missing ratings is equivalent to finding a complex relationship between each user and each item. Existing methods assumed this relationship as a fixed linear function, however this causes the predictions to be biased to the mean of ratings causing information loss. Therefore it is crucial to design a flexible model that reveals hidden, non-linear relationships between users and items.
In this paper, we propose NN-CMF (Neural Network Coupled Matrix Factorization), a neural network based method that predicts missing values of rating matrix by learning a non-linear function and latent matrices exploiting both rating matrix and auxiliary data. While conventional matrix factorization methods predict missing values through the inner product of latent vectors, NN-CMF learns a general non-linear function for latent vectors and therefore provides more accurate prediction. Experiments show that NN-CMF outperforms the conventional coupled matrix factorization methods by up to 5.4%. Our method is especially superior in improving accuracy on sparser datasets, making it more useful for real world applications.

Paper

NN-CMF is described in the following paper:

Code

The source codes used in the paper are available. [Download]

Datasets

Name# Users# Items# Auxiliary rows# Auxiliary columnsDensityDownload
MovieLens 100K9431682168219 Ratings: 6.30%, Auxiliary matrix: 100%,
Total: 8.15%
Link
MovieLens 1M60403706388319 Ratings: 4.47%, Auxiliary matrix: 100%,
Total: 4.77%
Link
FilmTrust15082071609732 Ratings: 1.14%, Auxiliary matrix: 0.42%,
Total: 1.04%
Link
Epinions401631397383396049288 Ratings: 0.01%, Auxiliary matrix: 0.03%,
Total: 0.02%
Link
Ciao (item coupled)17615161211612117 Ratings: 0.03%, Auxiliary matrix: 100%,
Total: 0.12%
Link
Ciao (user coupled)176151612114384299 Ratings: 0.03%, Auxiliary matrix: 0.65%,
Total: 0.04%
Link

People