Recurrent Event Networks Based on Subgraph and Attention Enhancement

Abstract

Temporal knowledge graph (TKG) reasoning, as an essential research direction in natural language processing, focuses on capturing the dynamic changes in entities and relationships over time. However, the inference task of predicting potential future events faces significant challenges, as it must deal with uncertainty, complexity, and missing data. To this end, this study proposes a new TKG extrapolation model SubRE-NET, based on the Recurrent Event Network(RE-NET). The model performs reasoning by aggregating local and global information, and introduces a subgraph in the encoding stage to enhance the ability to capture local correlations and temporal features of events within a time window. At the same time, the attention mechanism is introduced to solve the problem in which the original model cannot distinguish the importance of nodes and relationships, and the subgraph is further enhanced by cropping technology. In the aggregation stage, an extended relational graph convolutional network(RGCN) was adopted to overcome the limitations of the original model, which cannot capture temporal information when locally and globally aggregated. The experimental results show that, compared with the baseline model RE-NET, our SubRE-NET model achieved significant performance improvements on three event-based datasets and two public knowledge graph datasets, with an average MRR performance improvement of 11.49%. Simultaneously, the average performance of the Hits@1 metric is improved by 12.38%.

Keywords

• Attention network
• time knowledge graph
• subgraph
• temporal extrapolation