Performance Comparison of Collaborative Filtering Prediction Methods on Recommendation System

Abstract

Recommendation systems were introduced as the computer-based intelligent techniques to deal with the problem of information overload. Collaborative filtering is a simple recommendation algorithm that executes the similarity (neighborhoods) between items and then computes the missing data predictions. A serious limitation of collaborative filtering is the sparisity problem, referring to the situation where insufficient rating history is available for inferring reliable similarities. This research compares four prediction methods: Weighted Sum, Mean-Centering, Boosted Weighted Sum and Boosted Double Means Centering predictions. Boosting double means centering taken into account information of both users and items in order to overcome the potential decrease of accuracy due to sparsity when predicting the missing value. It tries to capture the user and item biases from the whole effects so as to enable the better concentrating on user-item interaction. Furthermore, ensemble learning will improve the performance collaborative filtering method because an ensemble of collaborative filtering models based on a single collaborative filtering algorithm considered the problem of sparsity, recommender error rate and sample weight update. Rating history in Book-Crossing dataset with 91% sparsity level is used to evaluate the missing rating predictions and the performance comparison of rating predictions on two traditional collaborative filtering and two boosting collaborative filtering frameworks. Experimental results shows that the proposed boosted double mean centering framework improve the prediction accuracy than the two traditional collaborative filtering and the other boosting prediction algorithm.