Multi-scale Community Detection in Subspace of Attribute

Abstract

The goal of attributed community detection is to search a partition of the network such that there is high cohesion within each group and low coupling between two groups. We argue that multiple partitions of the attributed network should be captured with different semantics and community detection should be approached from the perspective of attribute subspace. In this paper, we integrate spectral wavelets with attribute subspace, and develop a framework of Multi-scale Community Detection in Subspace of Attribute (MCDSA). Our idea is to implement graph partitioning via scale-dependent modularity and independent attribute subspaces, thus making our model more flexible and effective. In MCDSA, communities at each scale have independent attribute subspace, which is helpful to analyze the importance of each attribute under different network partition, better revealing the relationship between nodes. Extensive experiments on multiple benchmark datasets show that, the quality of community detection can be remarkably enhanced under the regime of attribute subspaces, achieving the state-of-the-art performance.