Regression-Based Fusion Prediction for Collaborative Filtering

Existing memory-based collaborative filtering techniques predict the unknown preference by taking weighted average of ratings by similar users on the active item or ratings of similar items by the active user, which just make use of the predictive power of ratings located in the row or column of user-item matrix corresponding to the active user or the active item. Due to sparsity, it is possible that some highly similar users have not rated the active item or some highly similar items have not been rated by the active user, resulting in they contributing nothing to the prediction. In this paper, first we propose improved regression-based methods to model the prediction of individual rating for the unknown preference. Then we propose regression-based fusion prediction (RBFP) algorithm, which adopts two-stage linear regression technique to exploit the predictive power of these unrated but highly similar items and these highly similar users who have not rated the active item. We have conducted extensive experiments, especially, we have investigated the sensitivity of parameters over the time, performance variation with the size of memory space available. Having done comparisons with some popular recommendation algorithms, we can conclude that our proposed methods can indeed improve the performance of collaborative filtering.