Attentive Hybrid Collaborative Filtering for Rating Conversion in Recommender Systems

Recommendation models that use collaborative filtering consider the influence of friends and neighbors when recommending suitable items for a target user. Most of these neighborhood-based models use the actual ratings from neighbors to predict the ratings of the target user toward target items, which often leads to a low accuracy prediction caused by the improper rating-range problem. Recently, rating conversion methods have been proposed to address this issue. Because each friend/neighbor can have a different level of influence on the target user, we propose a friend module, which converts their ratings to match the target user's perspective and assigns different weight to each user before modeling latent relations and predictions. In rating conversion, ratings that involve explicit feedback are important. Instead of the traditional approach to user embedding, we propose a novel approach that uses explicit feedback. This can express user features better than traditional methods and can then be used to convert ratings to match the target user's perspective. For better representation and recommendation, we also learn latent relations between each user and item by adopting knowledge graph ideas, which leads to more accurate results. The FilmTrust and MovieLens datasets are used in experiments comparing the proposed method with existing methods. This evaluation showed that our model is more accurate than existing methods.