Answers are grounded in this paper and cite their sources.
Existing resource recommendation methods primarily rely on knowledge graphs to model relationships between users and resources. However, evolving user preferences often degrade recommendation performance. In this article, we propose Temporal Preference-guided Knowledge Graph Recommendation (TPKG), which integrates temporal-aware user interest modeling (TAUI) with User-Preference-Guided Knowledge Graph Construction and Propagation, together with cross-view contrastive self-supervised learning. The central mechanism uses the same temporally inferred preference state to select relation types before propagation and to condition their attention logits. Our framework first infers dynamic user preferences from historical interactions through behavior-type-aware weighting, exponential temporal decay, and interest-diversity correction. It then constructs a personalized knowledge subgraph by filtering relations whose compatibility with the current preference state exceeds a predefined threshold, thereby restricting propagation to preference-compatible relation types. To enhance representation learning under sparse interaction conditions, temporal and knowledge-graph views are jointly optimized through intra-view and cross-view contrastive objectives, explicitly aligning behavioral sequence representations with knowledge graph-enhanced item representations. The contrastive objectives require no additional labels, while the framework uses an external knowledge graph and item-entity alignments. On four datasets, TPKG improves the reported mean Recall@10 over MDiffCL by 1.9–11.0%.
No takes yet. Share an insight, caveat, or question.
Cui et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: