New Loss Function Improves Heterogeneous Graph Recommendation Systems

Recommendation systems frequently rely on Graph Neural Networks trained on heterogeneous bipartite graphs, which often contain relationships of varying cardinalities, such as one-to-many user-item preferences and one-to-one user-attribute features. A common issue arises when a single loss function, like Bayesian Personalized Ranking (BPR), is applied universally. While BPR performs well for the core recommendation task, it can inadvertently cause attribute embeddings to degrade, making them nearly random. This "silent failure" often goes unnoticed by standard ranking metrics but contaminates user node embeddings, negatively impacting critical downstream applications like personalization and segmentation. To counter this, a new approach called Cardinality-Decomposed Loss (CDL) has been introduced. CDL integrates both Cross Entropy (CE) and BPR, allowing the model to collectively optimize for diverse relational cardinalities. The research confirms that CE and BPR can conflict within the shared encoder's parameter space. Evaluations across five datasets demonstrate that CDL consistently enhances the discriminability of attribute embeddings. It also shows that ranking performance (NDCG) improves when attributes provide meaningful preference signals, but can conflict when correlations are weak. A lambda parameter helps navigate this trade-off, revealing that dataset behavior is influenced by semantic alignment (attribute's predictive power for preferences) and topology leakage (whether graph connectivity already encodes preferences).