Community-oriented attributed network embedding

Abstract

Network embedding aims to map vertices in a complex network into a continuous low-dimensional vector space. Meanwhile, the original network structure and inherent properties must be preserved. Most of the existing methods merely focus on preserving local structural features of vertices, whereas they largely ignore the community patterns and rich attribute information. For example, the title of papers in an academic citation network could imply their research directions, which are potentially valuable in seeking more meaningful representations of these papers. In this paper, we propose a Community-oriented Attributed Network Embedding (COANE) framework, which can smoothly incorporate the community information and text contents of vertices into network embedding. We design a margin-based random walk procedure on the network coupled with flexible margins among communities, which limit the scope of random walks. Inspired by the analogy between vertex sequences and documents, the statistical topic model is adopted to extract community features in the network. Furthermore, COANE integrates textual semantics into representations through the topic model while preserving their structural correlations. Experiments on real-world networks indicate that our proposed method outperforms six state-of-the-art network embedding approaches on network visualization, vertex classification and link prediction.