Diversified Recommendation Using Graph Neural Networks Invariant to User-level Category Preference

Recommendation systems designed to address information overload increasingly need more diversity, significantly impacting user satisfaction. In this paper, we introduce two strategies to enhance diversity within these systems without drastically reducing accuracy. The first strategy involves using an invariant loss to category preference, effectively distinguishing between item-specific and category preferences while minimizing the latter's influence. The second strategy proposes a graph neural network-based learning sample selection process, mitigating the undue influence of specific categories on certain users. Utilizing a dataset from the web service Taobao, we quantitatively demonstrate the effectiveness of these methods, offering a balanced approach to maintaining both accuracy and diversity in recommendation systems.