Multi-layer Long Short-Term Memory Recursive Perceptual Network for Temporal Knowledge Graph Representation Learning

Siling Feng,Ziqi Ding,Mengxing Huang,Qian Liu

DOI: https://doi.org/10.1109/prai59366.2023.10331942

2023-01-01

Abstract:Temporal Knowledge Graphs (TKG) has been widely applied in many fields of life, and the reasoning of TKG has naturally become a research hotspot for many researchers. To set up extrapolation for TKG, we need to observe historical information, reflect entity preferences through historical information, and ultimately achieve the goal of predicting future events. However, for current inference models, how to make the model obtain more stable and accurate prediction results is still a challenge. Therefore, we propose a model called MLRPN (GCN-based Multi-layer Long Short-Term Memory Recursive Perceptual Network) that combines Graph Convolution Networks (GCN), Long Short-Term Memory (LSTM) and Multi-Layer Perceptrons (MLP) for regression modeling of temporal knowledge graph sequences. This model obtains evolutionary representations of entities and relations, and ultimately achieves prediction tasks for entities and relations. It consists of two modules: the Evolutionary Modeling Unit and the Score Calculation Function Unit. The Evolutionary Modeling Unit encodes and regresses the historical KG sequence to obtain evolutionary representations of entities and relations with longer temporal dependencies. The Score Calculation Function Unit contains the score functions for the two sub-tasks, which calculate the optimal solutions through score functions to obtain the best prediction output results. MLRPN was evaluated on four common TKG datasets, the experimental results showed that MLRPN outperforms other models on multiple datasets.