Temporal Knowledge Graph Embedding based on Multivariate Gaussian Process

Abstract

Recently, reasoning over Temporal Knowledge Graph (TKG), such as link prediction, has become an attractive research topic. Numerous Temporal Knowledge Graph Embedding (TKGE) methods have been proposed to map the entities and relations in TKG to the high-dimensional representations for further reasoning tasks. However, most existing TKGE methods which mainly based on deterministic vector embeddings, still have two drawbacks. On the one hand, they mainly model temporal evolution of entities and relations by a deterministic function of time, which captures the global trends but fails at the surging local fluctuations. On the other hand, they mainly focus on the semantic meaning of embeddings, while losing the sight of temporal uncertainties of the embeddings. To tackle such limitations, in this paper, we propose a novel approach to mapping the entities and relations in TKG to multivariate Gaussian Processes (MGP). With the flexibility and capacity of MGP, the global trends as well as the local fluctuations can be simultaneously modeled. Moreover, the temporal uncertainties can be also captured with the kernel function and covariance matrix of MGP. Experimental results show the effectiveness of the proposed approach on two real-world benchmark datasets compared with some state-of-the-art TKGE methods.

Tasks

Reproductions