An enhanced significance weighting approach for collaborative filtering

Collaborative filtering (CF) is a popular technique for rating prediction in recommender systems. CF tries to predict the user´s rating on an unseen item based on other similar users´ ratings. Computing similarity between users is dominantly carried out using correlation methods such as the Pearson correlation coefficient. These methods compute similarity only based on co-rated items. Due to sparsity of rating data, it is probable for similarity values to be computed based on only few co-rated items. These values do not necessarily reflect real users´ preferences. In other words, they are insignificant. As earlier studies suggest, the weight of these values should be decreased. In this paper, we show that it is insufficient to consider cases where there are only few co-rated items. We propose an enhanced approach, which modifies the similarity weights in all cases proportionally to the number of co-rated items. Experimental results show that our proposed approach substantially improves the prediction performance compared with previous studies. Parameter independency is another improvement of this approach, which makes it easy to use.