A Framework of Data Fusion Through Spatio-Temporal Knowledge Graph

Abstract

Data fusion aims to integrate data from different sources that represent the same real-world object from different views. The fusion of multi-source data can reduce the uncertainty and supplement the missing information of single-source knowledge. However, when data is generated from heterogeneous resources, it would be difficult to merge them straightforwardly by traditional data integration methods. So in this paper, we propose a knowledge graph based data fusion framework to integrate spatio-temporal information extracted from two different sources, i.e., travel notes and mobile phone data respectively, which record human travelling behaviors from different views. Firstly, we introduce a method of constructing a path-based spatio-temporal knowledge graph from different sources. Then, a long-path-based knowledge graph embedding is applied to learn entity representations of different knowledge graphs jointly, which can eliminate the heterogeneity of the information from different sources. An attenuation mechanism for modeling the long path relation is proposed in order to improve the representation learning for long path based knowledge graph. Finally, the entities are aligned considering context information to obtain the travelling knowledge in a unified and enhanced form. The experimental results show that compared with the baselines, the long-path-based knowledge graph embedding method is more suitable for the knowledge graph constructed in this paper. And through entity alignment, the two knowledge graphs can be fused to offer more information for subsequent applications.