Community detection is a significant methodology in network science. Traditional methods show limitations in dealing with multi-scale and high-dimensional complex data. As one of the most popular unsupervised algorithms, affinity propagation algorithm (AP) has been widely applied in community detection. However, its negative Euclidean similarity and inflexible parameter may lead to high time or memory consumption and excessive detection. Thus, this article presents a novel affinity propagation algorithm in t-distribution (APT), integrated with manifold learning, for detecting community structure. In APT algorithm, the data is compressed by dimensionality reduction, and joint probability is applied to construct the similarity matrix. Further, based on optimized modularity, parameters are adjusted to improve the accuracy. Experiments show that APT has better adaptability and universality than AP algorithm. In contrast to other mainstream algorithms, our algorithm can extract more meaningful communities from multi-scale and high-dimensional networks.

社区检测是网络科学中的一种重要方法。传统方法在处理多尺度和高维复杂数据方面显示出局限性。作为最流行的无监督算法之一，亲和力传播算法（AP）已广泛应用于社区检测。但是，其负欧氏相似度和不灵活的参数可能会导致较高的时间或内存消耗以及过多的检测。因此，本文提出了一种新颖的t分布亲和力传播算法（APT），该算法与流形学习相集成，用于检测社区结构。在APT算法中，通过降维压缩数据，并应用联合概率构造相似度矩阵。此外，基于优化的模块化，可以调整参数以提高准确性。实验表明，APT比AP算法具有更好的适应性和通用性。与其他主流算法相比，我们的算法可以从多维和高维网络中提取更有意义的社区。