GTR: An explainable Graph Topic-aware Recommender for scholarly document

Abstract

In the ever-expanding digital library of scholarly articles, navigating through the vast amount of available research papers to find relevant work poses a significant challenge to researchers. Addressing this issue, we introduce the Graph Topic-aware Recommender (GTR), an innovative end-to-end deep neural model tailored for scholarly recommendation systems. Unlike traditional methods that primarily rely on Collaborative Filtering, Content-Based Filtering, and Graph-Based approaches with limited consideration of the intricate citing logic within scientific documents, GTR captures the nuanced relationships and citing topics inherent in scholarly networks. By leveraging an advanced neural topic modeling technique, GTR transfers item-to-user recommendation into an item-to-item framework, facilitating a more accurate and contextually relevant paper recommendation process. Our study leverages the contextual richness of scholarly networks through Graph Neural Networks (GNNs), addressing the overlooked aspect of differentiated semantic relationships within these networks. The model stands out by effectively mining relation topics and conducting differentiated representations to minimize information redundancy, which enables GTR to adaptively infer latent citation topics, enhancing the model’s explainability and recommendation accuracy. Besides, the optimization function of GTR incorporates a novel component pooling module, designed to encode the sub-graph information of samples without traditional message passing, thereby improving the model’s efficiency and scalability. Through comprehensive experiments on multiple real-world scholarly datasets, GTR demonstrates superior performance over existing state-of-the-art models, offering both high accuracy and explainability in its recommendations.