Label Propagation-Based Membership Degree Computation towards a Computationally Efficient Fuzzy Community Detection Approach

This paper proposes a novel fuzzy community detection (FCD) approach, which we term as ‘Label Propagation-Based Fuzzy Community (LaProFC)’, and shows that it has the ability to outperform the existing FCD approaches. While designing the proposed FCD approach, we introduce a new compound type similarity metric termed ‘proportion of common neighbors and edges-based similarity (CCS)’ to compute similarity between two neighboring nodes. By executing local exploration on graphs with modified local random walk (mLRW), most similar neighbors of each node are identified; and based on the directions of most similar neighbors some tentative communities are generated. Afterward, these tentative communities are corrected and stabilized by iteratively computing membership degrees of each node using a novel label propagation-based membership computation function. We also propose a novel edge-density-based technique called ‘community-weight based tie-breaking (CTB)’, which is incorporated with the membership degree computation function. We conduct extensive experiments with both real-life and synthetic datasets and show the working of the proposed approach. Our Proposed LaProFC approach outperforms baseline approaches in terms of popular quality and accuracy metrices including modularity and normalized mutual information. Further, popular multi-criteria decision making (MCDM) tools are used to show supremacy of the proposed approach by computing the ranks of different approaches through two sets of accuracy and quality metrices. Our proposed LaProFC approach supersedes other approaches in terms of faster computations and asymptotic time complexity.