Item Enhanced Diversification in Recommendation Systems Using Graph Neural Networks

It remains a challenging problem of how to achieve the tradeoff between accuracy or diversity, especially in high-dimensional item spaces. In this paper, we put forth an Item Enhanced Diversification Framework based on Graph Neural Networks (GNNs) to tackle this problem. The model makes recommender systems more diverse while still considering novelty. It uses a sparse graph encoding and lightweight inference modules for scalability and cold-start robustness. Extensive experiments on multiple datasets show the significant performance improvement in diversity metrics as well as user satisfaction, indicating the effectiveness of the proposed framework.