A generative and discriminative model for diversity-promoting recommendation

Diversity-promoting recommender systems with the goal of recommending diverse and relevant results to users, have received significant attention. However, current studies often face a trade-off: they either recommend highly accurate but homogeneous items or boost diversity at the cost of relevance, making it challenging for users to find truly satisfying recommendations that meet both their obvious and potential needs. To overcome this competitive trade-off, we introduce a unified framework that simultaneously leverages a discriminative model and a generative model. This approach allows us to adjust the focus of learning dynamically. Specifically, our framework uses Variational Graph Auto-Encoders to enhance the diversity of recommendations, while Graph Convolution Networks are employed to ensure high accuracy in predicting user preferences. This dual focus enables our system to deliver recommendations that are both diverse and closely aligned with user interests. Inspired by the quality vs. diversity decomposition of Determinantal Point Process (DPP) kernel, we design the DPP likelihood-based loss function as the joint modeling loss. Extensive experiments on three real-world datasets, demonstrating that the unified framework goes beyond quality-diversity trade-off, i.e., instead of sacrificing accuracy for promoting diversity, the joint modeling actually boosts both metrics.