Relation-Entity Hybrid Learning Graph Model for Few-Shot Temporal Knowledge Graph Forecasting

Shiqi Fan, Hongyi Nie, Ruibing Wang, Quanming Yao, Haotong Du, Yang Liu, Zhen Wang

Abstract: Recent research has widely substantiated the exceptional forecasting capabilities of Temporal Knowledge Graphs (TKGs). However, a challenging real-world problem is that over time, new entities continually emerge, often possessing very limited historical information, making model training considerably difficult. On the one hand, most existing models primarily focus on addressing new entity problems in static KG, disregarding the significance of temporal information. On the other hand, models optimized for TKG often overlook the logical correlations between relations, placing greater emphasis on mining interactions between entities. In this paper, we propose a novel Relation-Entity Hybrid Learning Graph Model to address this challenge. The model comprises three key components: (1) we construct a Relation Enhance Module, aiming to explore the logical correlations between relations to enhance relation embeddings; (2) we design an Entity Learning Module, which encodes entity-related graph structures based on relation embeddings; (3) we employ a carefully designed Time Embedding Module to capture short-term and periodic temporal information. To comprehensively evaluate the model’s performance, we introduce three new TKG few-shot forecasting datasets. Through extensive experiments, our approach exhibits significant advantages on these three datasets, surpassing baseline models.