Cross-feature interactive temporal knowledge graph reasoning with evolving retention mechanism

Abstract

Temporal knowledge graph (TKG) reasoning emphasizes deducing absent connections within evolving knowledge graphs (KGs), which is essential for comprehending dynamic engineering informatics. However, the ongoing dynamic evolution of TKGs presents significant challenges for accurate predictions. To address this challenge, this paper proposes a cross-feature temporal evolution network (CFTENet), which designs an evolving retention mechanism establishing a knowledge forgetting threshold to lock in snapshots of continuous evolution. The importance of knowledge gradually diminishes until the information becomes outdated and is completely forgotten. Historical information at previous time points is preserved in current snapshot to simulate continuous dynamic evolution of knowledge. Moreover, CFTENet incorporates a cross-feature interaction module, leveraging a multilayer dilated convolutional network and a residual network to grasp cross-feature intricate interactions among and across entity and relation characteristics. The proposed model improves the reasoning ability and resilience to unseen data. Comprehensive testing on four benchmark datasets (ICEWS14, ICEWS18, GDELT, WIKI) demonstrates that our model achieves significant performance improvements, surpassing the baseline methods by 1.5 %, 8.8 %, 6.5 %, and 2.2 %, which highlights its effectiveness in TKG reasoning.