DynaKG: Dynamic Knowledge Graph Attention With Learnable Temporal Decay for Recommendation

Abstract

Knowledge graph-based recommendation systems excel at capturing semantic relationships but struggle to model temporal dynamics in evolving user preferences, which is a critical limitation for real-world applications where user behaviors and item popularity change continuously over time. The core challenge lies in the existing approaches that treat knowledge graphs as static structures and fail to capture the temporal evolution, sequential dependencies, and scalability demands of dynamic systems. We propose a Dynamic Knowledge Graph attention network (DynaKG) that addresses these limitations through three key innovations: 1) learnable temporal decay mechanisms with context-specific parameters that adapt to different node and edge types, 2) hybrid Fourier-positional temporal encoding that captures both periodic and aperiodic patterns, and 3) time-aware contrastive learning with incremental updates that reduces computational overhead by 34% while maintaining quality. We provide a theoretical analysis of the convergence properties of the temporal attention mechanism and empirically demonstrate that DynaKG substantially outperforms state-of-the-art baselines across three datasets: 8.0% improvement in Recall@20 on Yelp, 9.0% on Amazon-Book, and 7.0% on MovieLens-10M, while achieving 3.6 times parameter efficiency and providing interpretable temporal explanations.