Research on Temporal Knowledge Graph Reasoning Based on Trend and Dynamic Change Perception Enhancement

Temporal knowledge graphs (TKGs), as an effective means of modeling dynamic relationships among entities, have shown great potential in tasks such as event prediction in recent years. However, most existing reasoning approaches tend to overlook the diversity and complexity of historical information. In reality, reasoning at the current timestamp is often constrained by the limited scope of historical data and the influence of unobserved latent factors. There are three major limitations in current TKG reasoning methods: (1) the inability to effectively highlight the importance of historical snapshots relevant to the current query when integrating both local and global history; (2) the neglect of temporal trends inherent in the evolution of facts, resulting in insufficient modeling of evolutionary patterns; and (3) the lack of effective mechanisms for capturing abrupt, short-term changes in the temporal dimension of facts. To address these challenges, we propose a novel Trend- and Variation-aware Contrastive Learning Network. Specifically, we introduce a local-global contrastive learning mechanism to guide the model's focus toward historical information that is more relevant to the current query. We further design a trend-aware attention module to capture the regularities and temporal evolution patterns in long-term entity representations. Additionally, a time-aware convolution module is developed to perceive abrupt dynamic changes in entity states across consecutive time slices, enabling the model to better integrate this information with current context representations. Experimental results on four benchmark TKG datasets demonstrate that our model outperforms several state-of-the-art baseline models in prediction tasks, showcasing its superior generalization ability and effectiveness in modeling complex temporal evolution patterns.