基于标记注意力机制的社区发现算法

针对网络的聚类进行研究，提出了一种基于标记注意力机制的社区发现算法，网络特征通过标记节点频率及反示例节点频率联合度量，为使网络特征的度量更加关注于示例节点的细节信息，引入注意力机制来处理网络特征。社区划分由复杂网络预处理、网络节点的策略、社区博弈归并三个部分组成，其中网络节点的策略由无贡献节点归并、节点到社区的判断以及节点逻辑标记和的判断三个步骤组成。实验借助于真实网络进行验证，在归一化互信息、模块度、社区划分数量及运行时间四个方面，基于标记注意力机制的社区发现算法都优于其它社区发现算法。在实际生活中应用此算法，能够更加直观地显示网络内部之间存在的联系。

Community discovery algorithm based on label attention mechanism

For network clustering, this paper proposes a community discovery algorithm based on label attention mechanism. Network features are jointly measured by label node frequency and inverse example node frequency, and attention mechanism is introduced to handle network features in order to make the metric of network features more focused on the detailed information of example nodes. Community division is composed of three parts: complex network preprocessing, network node strategy, and community game merging. The network nodes strategy is consists of three steps: the non-contributing node merging, the judgment of the node to the community and the judgment of the degree of the node. The experiments are verified with the help of real networks. The results show that the community discovery algorithm based on label attention mechanism outperforms other community discovery algorithms in four aspects: normalized mutual information, modularity, number of community divisions and running time. Applying this algorithm in real life can more intuitively show the existing connections within the network.