Historical trends and normalizing flow for One-Shot temporal knowledge graph reasoning

Abstract

Temporal Knowledge Graphs (TKGs) have garnered significant scholarly interest for their ability to dynamically represent the evolution of real-world events. However, the presence of the long tail effect in knowledge graphs necessitates the exploration of one-shot temporal knowledge graph reasoning. Concurrently, the issues of out-of-distribution and overfitting heavily impact the performance of models. Empirical evidence from real-world cases has demonstrated the significance of historical trend information in accurately predicting future events, and normalizing flows effectively address out-of-distribution and overfitting challenges. Thus, this paper proposes a novel approach named Historical Trends and Normalizing Flow for One-shot Temporal Knowledge Graph Reasoning (HNOT). Specifically, HNOT combine normalizing flows with a historical information aggregator to obtain more complex distributions, thereby resolving out-of-distribution and overfitting problems and capturing global representations of TKGs. In contrast to existing models that focus solely on entity and relationship modeling, we have introduced a novel historical trend information aggregator. This aggregator is designed to extract trend information from temporal knowledge subgraphs. By employing trend information as the primary modeling object, it facilitates the exploration of trend variations at different timestamps. This approach serves as an auxiliary source of global information. Subsequently, we integrate these two distinct types of information to achieve predictions of future events. Experimental results demonstrate the superiority of the proposed method over state-of-the-art baselines on three benchmark datasets.