Contextual Recommendations: Dynamic Graph Attention Networks With Edge Adaptation

Abstract

Recommender systems have witnessed a great shift in leveraging contextual information as an auxiliary resource to improve the quality of the recommendations. These recommendation problems are addressed as link prediction tasks within bipartite graphs, where user and item nodes are connected by edges labeled with binary values or rating information. This paper introduces a new architecture: Dynamic Graph Attention Network with Adaptive Edge Attributes (DGAT-AEA). Comprising multiple layers of dynamic Graph Attention Networks, designed to efficiently handle contextual recommendations. Our method is distinguished by its ability to update user and item representations while adapting the attributes of the connections between them during learning. This enables the capture of complex relationships within user-item interactions using contextual information. By optimizing model parameters and adapting edge features according to the user-item-context relationship, our approach outperforms existing methods regarding recommendation accuracy and novelty, as demonstrated by experiments on benchmark datasets.