چكيده لاتين
Given the role of social networks in daily life and the vast amount of data generated within these networks, analyzing such large-scale data is of great importance. One of the key operations in this analysis is community detection, which aims to identify groups of nodes with dense intra-group connections and sparse inter-group connections, providing valuable insights into the structure, behavior, and dynamics of networks. Communities in social networks are classified into two categories: disjoint communities, where each network member belongs to only one community, and overlapping communities, where individuals can simultaneously participate in multiple communities. Identifying overlapping communities is crucial given real-world scenarios. Moreover, most real-world networks contain relationships that indicate friendship or hostility, trust or distrust, which are referred to as signed networks. Community detection in these networks faces the challenge of considering the signs of links, making the identification of overlapping communities even more complex. This study aims to simultaneously address the challenges of signed networks and overlapping community detection. While several algorithms have been proposed in this field, existing methods still face challenges, the most significant being their high computational complexity as network size increases. One of the well-known community detection algorithms is the Label Propagation Algorithm (LPA), which benefits from low computational complexity, ease of implementation and use, and no requirement for prior knowledge about communities. Although this algorithm has been developed for detecting overlapping communities in unsigned networks and disjoint communities in signed networks, its application for detecting overlapping communities in signed networks remains unexplored. In this study, LPA is extended using multiple approaches to detect overlapping communities in signed networks. These approaches are categorized into two groups: those that utilize network weighting and those that do not. Weight-based approaches also address the randomness issue of LPAʹs community detection results. The proposed methods were evaluated based on modularity and frustration criteria, leading to the selection of the most suitable approach. The optimal values obtained from its execution for modularity and frustration on the Epinions dataset were 0.38 and 0.07, on the Slashdot dataset were 0.43 and 0.21, and on the Bitcoin dataset were 0.31 and 0.06. Comparing these values with the optimal results from two LPA-based approaches used for detecting overlapping communities in unsigned networks indicates that the proposed approach achieves better performance across all three datasets. Therefore, incorporating edge signs improves community detection performance and yields superior results. Additionally, the proposed approach, which employs network weighting and structural balance measures, offers acceptable computational complexity compared to existing methods while leveraging the advantages of LPA.
