Feature Interaction for Temporal Knowledge Graph Extrapolation

The emergence of knowledge graphs has sparked interest in temporal knowledge graph reasoning, which incorporates temporal data into static graphs. Recently, significant advancements in TKG extrapolation have focused on predicting future events using historical data. However, many existing methods overlook the complex dynamics of entity and relation interactions over time. To address this, we developed a novel method called FIM for temporal knowledge graph reasoning. FIM enhances interaction modeling by identifying interaction pairs through feature permutation, reinforcing interactions with chequer reshaping, using a refined-SENet with a gate mechanism for calibration, and utilizing circular convolution for boundary information. Rigorous experiments and comparative analysis demonstrate FIM’s exceptional performance in link prediction tasks.