Attention-based recommender systems by applying region embedding

Abstract

Recommender system is a system that is used to search for an item that a user prefers and recommends it to another user. The Recommender Systems often uses the consistency with previous usage history of the target user and other users in the system or the user’s preferences to predict scores. However, they still lack consideration of latent relations in the sequence around a target user and a target item. Hence, this research proposes a new recommendation method that can extract a latent relation in a historical sequence of a target user and a target item by applying Region Embedding with the Local Context Unit in order to utilize that latent relation for personalized rating predictions. However, this method does not serve for personalized recommendation for each target user. Therefore, this research also proposes a specific set of neighbors to different target users in order to differentiate among target users even if having the same set of neighbors by applying method of Attention mechanism. This method calculates Attention scores which are similar to the similarity score between neighbors and target users. Therefore, the proposed method can predict rating scores which are different for different target users. The dataset used in the experiment is MovieLens. This research compares the efficiency of the proposed method with Neural Collaborative Filtering recommendation. Moreover, the experimental results between our method with Attention was compared to our method without Attention. Finally, the models were evaluated with NDCG@K and HitRate@K. The results show that our model is better than NCF. Moreover, our model with Attention is better than our model without Attention.