Dynamic Item Weighting and Selection for Collaborative Filtering

User-to-user correlation is a fundamental component of Collaborative Filtering (CF) recommender systems. In user-to-user correlation the importance assigned to each single item rating can be adapted by using item dependent weights. In CF, the item ratings used to make a prediction play the role of features in classical instance-based learning. This paper focuses on item weighting and item selection methods aimed at improving the recommendation accuracy by tuning the user-to-user correlation metric. In fact, item selection is a complex problem in CF, as standard feature selection methods cannot be applied. The huge amount of features/items and the extreme sparsity of data make common feature selection techniques not effective for CF systems. In this paper we introduce methods aimed at overcoming these problems. The proposed methods are based on the idea of dynamically selecting the highest weighted items, which appear in the user profiles of the active and neighbor users, and to use only them in the rating prediction. We have compared these methods using a range of error measures and we show that the proposed dynamic item selection performs better than standard item weighting and can significantly improve the recommendation accuracy.