Abstract Because of its broad real-life application, community detection (in the realm of a complex network) is an attractive challenge to many researchers. However, current methods fail to reveal the full community structure and its formation process. Thus, here we present SIMGT, an effective and Scalable approach that detects overlapping communities: Integrating social identity Model and Game Theory. Inspired by social identity theory and nodes’ high-order proximities, first we weight and rewire the original network, then we associate each node with a new utility function. Next, we model community formation as a non-cooperative game among all nodes, and we regard the nodes as self-interested players. Further, we use a stochastic gradient-ascent method to update players’ strategies toward different communities, and prove that our game greatly resembles and matches how a potential game works (in the classical sense in game theory), indicating that the Nash equilibrium point must exist. Finally, we implement comprehensive experiments on several synthetic and real-life networks. The results show that whatever weighting strategy we choose, SIMGT can gain better performance on community detection task. In particular, SIMGT achieves a best result when we choose the Jaccard coefficient. After comparing SIMGT with six benchmark algorithms, we obtain convincing results in terms of how well the algorithms reveal communities, as well as algorithms’ scalability.

中文翻译：

---

一种有效且可扩展的重叠社区检测方法：整合社会身份模型和博弈论

摘要 由于其广泛的现实应用，社区检测（在复杂网络领域）对许多研究人员来说是一个有吸引力的挑战。然而，目前的方法未能揭示完整的社区结构及其形成过程。因此，我们在这里介绍 SIMGT，这是一种检测重叠社区的有效且可扩展的方法：整合社会身份模型和博弈论。受社会认同理论和节点的高阶邻近性的启发，我们首先对原始网络进行加权和重新布线，然后将每个节点与一个新的效用函数关联起来。接下来，我们将社区形成建模为所有节点之间的非合作博弈，我们将节点视为自利玩家。此外，我们使用随机梯度上升方法来更新玩家对不同社区的策略，并证明我们的博弈与潜在博弈的运作方式非常相似和匹配（在博弈论的经典意义上），表明纳什均衡点必须存在。最后，我们在几个合成和现实生活中的网络上进行了全面的实验。结果表明，无论我们选择何种加权策略，SIMGT 都能在社区检测任务上获得更好的性能。特别是当我们选择 Jaccard 系数时，SIMGT 取得了最好的结果。在将 SIMGT 与六种基准算法进行比较后，我们在算法揭示社区的程度以及算法的可扩展性方面获得了令人信服的结果。我们在几个合成网络和真实网络上进行了全面的实验。结果表明，无论我们选择何种加权策略，SIMGT 都能在社区检测任务上获得更好的性能。特别是当我们选择 Jaccard 系数时，SIMGT 取得了最好的结果。在将 SIMGT 与六种基准算法进行比较后，我们在算法揭示社区的程度以及算法的可扩展性方面获得了令人信服的结果。我们在几个合成和现实生活中的网络上进行了全面的实验。结果表明，无论我们选择何种加权策略，SIMGT 都能在社区检测任务上获得更好的性能。特别是当我们选择 Jaccard 系数时，SIMGT 取得了最好的结果。在将 SIMGT 与六种基准算法进行比较后，我们在算法揭示社区的程度以及算法的可扩展性方面获得了令人信服的结果。