A nonnegative Gumbel-based encoder–decoder approach for community detection

Graph clustering or community detection is normally used to comprehend the structure of complex networks and to retrieve meaningful information from network data. One of the most popular methods for unveiling community structure is Latent factor models which include mixed membership block model and nonnegative symmetric matrix factorization. These models generally focus on identifying a representation of a network which is distributed and low dimensional, that uncovers the community membership of nodes and apprehends its structural regularity. Existing methods are primarily dependent on recreating a network from a node’s representation such as encoder-decoder method, while restraining the representation to possess specific characteristics. However, these approaches do not use any selection and extraction technique to acquire the membership of nodes. Therefore, they fail to explain the meaning of community memberships and encounter computational issues. This research work proposes a Nonnegative Gumbel-based Encoder-Decoder (NGED) method to detect community structures in complex networks. By integrating Gumbel softmax with an encoder and a decoder, the proposed NGED method provides promising results than the contemporary methods for community detection. Moreover, it also captures the community meaning and handles the computational problems efficiently. We demonstrated efficacy of the proposed NGED method using various real-world datasets and the results of the detailed experiments indicate that the proposed technique gives better community structures.