Incorporating Similarity and Trust for Collaborative Filtering

Currently, most recommender systems are using collaborative filtering (CF) techniques. The main idea is to suggest new relevant items for an active user based on the judgements from other members in the like-minded community. However, these CF-based methods encounter the obstacles, such as sparse data, cold-start and robustness. This paper proposes to deal with these issues by associating similarity measurement from users’rating patterns with trust metric. After investigating the large data set from Epinions.com, we find that user similarity and trust are strongly correlated. This fact also explains why using trust (instead of user similarity) could lead to very close mean prediction accuracy in a Pearson Correlation Coefficient-like recommendation algorithm. Our novel method incorporates these two factors into one unified recommendation algorithm. The experimental results indicate that a good prediction strategy can come from filtering the ratings from the users who have high trust and low similarity or vice versa.