Time-sequence attention-enhanced graph convolution network for temporal knowledge graph inference

This paper introduces a novel Temporal Attention-Enhanced Graph Convolutional Network (TA-GCN) for temporal knowledge graph (TKG) inference, focusing on predicting missing facts based on historical data. We construct a time-sequence interconnection graph to capture relationships between static knowledge graph snapshots at different timestamps. A hierarchical attention-enhanced graph convolution network is then designed to effectively capture both short-term local and long-term global dependencies. Experimental results on four widely-used datasets show that TAGCN outperforms existing methods, demonstrating its effectiveness and superiority.