Mining Significant Granular Association Rules for Diverse Recommendation

Granular association rule is a new technique to build recommender systems. The quality of a rule is often evaluated by the confidence measure, namely the probability that users purchase or rate certain items. Unfortunately, the confidence-based approach tends to suggest popular items to users, and novel patterns are often ignored. In this paper, we propose to mine significant granular association rules for diverse and novel recommendation. Generally, a rule is significant if the recommended items favor respective users more than others; while a recommender is diverse if it recommends different items to different users. We define two sets of measures to evaluate the quality of a rule as well as a recommender. Then we propose a significance-based approach seeking top-k significant rules for each user. Results on the MovieLens dataset show that the new approach provides more significant and diverse recommendations than the confidence-based one.