Gaussian mixture embedding of multiple node roles in networks

• Published in: World wide web (Bussum) Vol. 23; no. 2; pp. 927 - 950
• Main Authors: Chen, Yujun, Pu, Juhua, Liu, Xingwu, Zhang, Xiangliang
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.03.2020; Springer Nature B.V
• Subjects: Clustering; Computer Science; Database Management; Embedding; Gaussian distribution; Information Systems Applications (incl.Internet); Network analysis; Nodes; Normal distribution; Operating Systems; Special Issue on Graph Data Management in Online Social Networks; Topology; Energy based learning; Network embedding; Gaussian mixture distribution; Graph mining
• ISSN: 1386-145X; 1573-1413
• DOI: 10.1007/s11280-019-00743-4

Network embedding is a classical topic in network analysis. Current network embedding methods mostly focus on deterministic embedding, which maps each node as a low-dimensional vector. Thus, the network uncertainty and the possible multiple roles of nodes cannot be well expressed. In this paper, we propose to embed a single node as a mixture of Gaussian distribution in a low-dimensional space. Each Gaussian component corresponds to a latent role that the node plays. The proposed approach thus can characterize network nodes in a comprehensive representation, especially bridging nodes, which are relevant to different communities. Experiments on real-world network benchmarks demonstrate the effectiveness of our approach, outperforming the state-of-the-art network embedding methods. Also, we demonstrate that the number of components learned for each node is highly related to its topology features, such as node degree, centrality and clustering coefficient.