WTNet: temporal knowledge graph reasoning network with windowing strategy and transformer-based fusion

Temporal knowledge graph (TKG) reasoning involves predicting future facts by identifying dynamic evolutionary patterns from historical data. The foundation of effective reasoning lies in the comprehensive identification and analysis of the temporal regularities inherent in the given facts. However, existing approaches often struggle with a trade-off: modeling isolated historical snapshots leads to information loss in long-term dependencies, while directly aggregating entire subgraphs introduces excessive noise and redundancy. To address these issues, we propose a novel neural network, named WTNet, with Windowing strategy and Transformer-based fusion for TKG Reasoning. WTNet uses a sliding temporal window to convert sequences of historical subgraphs into a compact set of temporal window graphs. These window graphs preserve temporal interactions across subgraphs while filtering out noise, thereby effectively capturing local temporal dynamics while minimizing irrelevant information. To further extract global temporal features across window graphs, we employ a transformer encoder to model long-range temporal dependencies. We conduct extensive experiments on four public TKG datasets, i.e ., ICEWS18, GDELT, YAGO, and WIKI. The experimental results show that WTNet achieves competitive performance compared with representative temporal knowledge graph reasoning models, reaching a peak mean reciprocal rank (MRR) of 83.42% on the WIKI dataset. Notably, statistical analysis confirms the model’s robustness, reducing performance volatility by over 50% compared to state-of-the-art baselines across varying historical lengths.