Multi-view Based Entity Frequency-Aware Graph Neural Network for Temporal Knowledge Graph Link Prediction

Abstract

Inferring missing facts in temporal knowledge graph (TKG) is a fundamental and challenging task. Existing models typically use representation learning to solve this problem. However, most of these models fall short of capturing multi-hop structural information and general preferences for future emerging facts when implementing the representation of target nodes. In addition, most of them use recurrent neural networks to achieve the aggregation of temporal information, which is not only less scalable as the time step increases, but also fails to explicitly address the problem of temporal sparsity of entity distribution in TKG. To address the above problems, we present a MEFGNN (Multi-view based Entity Frequency-aware Graph Neural Network) framework that learns node embedding to capture structural evolution of TKG by combining Multi-view Graph Neural Network (MGNN) and Entity Frequency-aware Attention Network (EFAN). Experiments on three real datasets show that MEFGNN outperforms state-of-the-art methods, our ablation study also validates the effectiveness of MGNN and EFAN.