TKGR-GPRSCL: Enhance Temporal Knowledge Graph Reasoning with Graph Structure-Aware Path Representation and Supervised Contrastive Learning

Lizhu Xiong, Yongpan Sheng, Lirong He

Abstract: Temporal knowledge graph reasoning (TKGR) aims to infer new temporal fact knowledge based on existing ones, which has become a major concern due to its wide demands, especially for time-sensitive modeling scenarios. However, existing temporal knowledge graph representation learning models usually have challenges in capturing complex graph structure-aware information and discriminating different relation expressions for unknown query prediction, which potentially limits the model performance on the TKGR task. In this paper, we propose a novel TKGR model named TKGR-GPRSCL. Specifically, we first introduce a maximum entropy-based random walk sampler, which can sample sufficient paths containing complicated graph structure-aware information in the TKG in form of path embedding. Then, we model the preference of different paths to the target entity, and the correlations between the target entity and its temporal property depending on transformations in the complex plane for encoding temporal fact. Finally, we advance supervised contrastive learning of relation patterns in light of the discrimination of different relation expressions to be preferably leveraged to predict entities for testing the queries, for those unknown queries in particular. We evaluate TKGR-GPRSCL on three TKGR benchmark datasets, and experimental results demonstrate that our model achieves superior performance compared to competitive baselines  (All the data and codes are publicly available at https://github.com/shengyp/TKGR-GPRSCL).

External IDs:dblp:conf/nlpcc/XiongSH24