A Multi-criteria Collaborative Filtering Recommender System Using Learning-to-Rank and Rank Aggregation

Recommender system suggests a top-N list from unseen items for its users through a prediction or a ranking order process. From the recommendation perspective, the item’s order in the generated list is more important than its predicted rating. Moreover, finding the top-N list for a multi-criteria recommendation is a challenging problem as we have many criterions for each item. One can find the average over all criteria; however, this requires a score from each criterion and hence a compensation effect will occur. This resembles many prediction-based recommendation systems working in parallel. Alternately, this paper proposes a three-step hybrid ranking order system for finding the top-N list for the multi-criteria recommendation system. The first step decomposes the multi-criteria user-item matrix into many single-rating user-item matrices while the second step finds partial-ranked lists for each item using a learning-to-rank method. This allows us to reflect the interest of the user for each criterion and then pass on this information for the next stage. The last step aggregates the partial-ranked lists into a global-ranked list using a ranking aggregation method. This will reduce the processing time and improve the recommendation quality by representing the user preference for each criterion. Three different sets of experiments are conducted on Yahoo!Movie dataset, and the results show that the proposed multi-criteria-ranking approach outperforms both the traditional no-ranking item-based collaborative recommendation and single-criteria-ranking approach that uses two popular learning-to-rank methods.