EvoREG: Evolutional Modeling with Relation-Entity Dual-Guidance for Temporal Knowledge Graph Reasoning

Temporal knowledge graph (TKG) reasoning, which aims to predict unknown future facts based on real-world historical information, has become a current research hotspot. Typically, TKG reasoning approaches mainly concentrate on modeling interactions among entities, while the abundant semantic interactions among relations are almost neglected. Besides, TKG reasoning approaches mostly utilize a single graph convolutional network (GCN) to learn representations for all snapshots, ignoring their differences in topological characteristics. In this paper, we propose a novel method, named EvoREG, to tackle these issues. Specifically, to fully explore the semantic interactions among relations, EvoREG constructs a hyper-relation graph corresponding to each snapshot, employing GCN with evolution mechanism to capture latent interactions contained in relations. Furthermore, to capture the dynamic changes of snapshots, EvoREG exploits recurrent neural networks (RNNs) to evolve the parameters of GCNs, fully mining the dynamic evolution of TKGs. Extensive experiments on widely-used benchmarks demonstrate the effectiveness of EvoREG for TKG reasoning. Our codes are available at https://github.com/scu-kdde/HGA-EvoREG-2024 .