Confidence on Collaborative Filtering and Trust-Based Recommendations

Memory-based collaborative filtering systems predict items ratings for a particular user based on an aggregation of the ratings previously given by other users. Most systems focus on prediction accuracy, through MAE or RMSE metrics. However end users have seldom feedback on this accuracy. In this paper, we propose confidence on predictions in order to depict the belief from the system on the pertinence of those predictions. This confidence can be returned to the end user in order to ease his/her final choice or used by the system in order to make new predictions. It takes into account some characteristics on the aggregated ratings, such as number, homogeneity and freshness of ratings as well as users weight. We present an evaluation of such a confidence by applying it on different collaborative filtering systems of the literature using two datasets with different characteristics.