The effect of rating variance on personalized recommendation

Recommender systems have made significant progress over the last decade and several industrial-strength systems have been developed. Typically, recommender systems try to predict people´s preferences and use accuracy indices such as mean absolute error to judge the performance of the algorithms. Recently, the diversity index is widely accepted as another metric. However, the ability of a recommendation algorithm to gain both accuracy and diversity at the same time remains largely unexplored. Variance based collaborative filtering is proposed as an improvement to satisfy the two measures. In this paper, we give a detail discussion on the effect of rating variance of collaborative filtering (CF) and get some different results. We find that the variance-based algorithm has limited ability to overcome the accuracy-diversity tradeoff. In fact the improvement in diversity comes at the expense of the precision. How to solve the dilemma is still a problem deserved further researching.