Abstract

Community detection is a crucial research orientation in complex network owing to its practical applications. Recently, the convolutional neural network (CNN) based edge classification community detection approach achieved impressive performance as a representative graph embedding method. However, existing methods rely heavily on the manually defined relationship between nodes and local feature representations of the edges, leading to the potential limitation of feature representation in complex real-world networks. To tackle this issue, in this paper, we propose a novel multi-feature fusion network for community detection, namely MFF-Net, which can embed edges of the network into comprehensive feature representations according to the intrinsic property of the network. Specifically, instead of utilizing local features of the edges only, we first propose to simultaneously consider local and non-local relationships of the edges through the local neighbors of the nodes and random node sequences sampled from a customized random walk. Then, we propose to convert local and non-local feature representations of the edges into grayscale images for edge classification via introducing a quantitative relationship between nodes. More importantly, those obtained feature representations are fused with an implicit manner in a latent space, which can leverage more comprehensive relationships for complex real-world networks. Extensive experiments on the computer-generated networks, small-scale and large-scale real-world networks, show that the proposed MFF-Net can achieve better performance compared with existing methods, in terms of quantitative and visualized results.

Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call