The central problem of a recommender system is how to represent the user and item embedding. To explore the user-item interaction, one of the traditional methods is Collaborative Filtering(CF), which learns user and item embeddings based on historical user-item interactions. However, the performance of CF is limited due to the sparseness of user-item interaction, also CF methods ignore the user’s social connection, which also affect user preferences. Social recommendation leverages social connection to alleviate data sparsity in the recommender system. We argue that attention mechanisms in Graph Neural Networks (GNNs) can effectively incorporate social information to characterize the social property of the users and improve the performance of recommender systems.
In this paper, we discuss cost-efficiency of deep relation-enhanced graph attention networks for social recommendation systems. In particular, we present our experiments on DREANRec to discuss whether having the attention mechanism in this module gives better cost-performance. Extensive experiments were implemented to prove the effectiveness of our approach.