一种基于因子图模型的半监督社区发现方法

A Semi-supervised Community Detection Method Based on Factor Graph Model

社区发现是社交网络分析中一个重要的研究方向.当前大部分的研究都聚焦在自动社区发现问题,但是在具有数据缺失或噪声的网络中,自动社区发现算法的性能会随着噪声数据的增加而迅速下降.通过在社区发现中融合先验信息,进行半监督的社区发现,有望为解决上述挑战提供一条可行的途径.本文基于因子图模型,通过融入先验信息到一个统一的概率框架中,提出了一种基于因子图模型的半监督社区发现方法,研究具有用户引导情况下的社交网络社区发现问题.在三个真实的社交网络数据(Zachary社会关系网、海豚社会网和DBLP协作网)上进行实验,证明通过融入先验信息可以有效地提高社区发现的精度,且将我们的方法与一种最新的半监督社区发现方法 (半监督Spin-Glass模型)进行对比,在三个数据集中F-measure平均提升了6.34%、16.36%和12.13%.

Community detection is an important research direction of social network analysis. Most of the current studies focused on automated community detection. However, in networks having missing data or noise, the ability for an automated community detection algorithm to discover true community structures may degrade rapidly with the increase of noise. On the other hand, semi-supervised community detection provides a feasible way for solving the above problem by incorporating priori information into the community detection process. In this paper, based on the factor graph model,by incorporating the priori information into a unified probabilistic framework, we propose a factor graph-based semisupervised community detection method. We evaluate the method with three different genres of real datasets（Zachary,Dolphins and DBLP）. Experiments indicate that incorporating priori information into the community detection process can improve the prediction accuracy significantly. Compared with a latest semi-supervised community detection algorithm（semi-supervised spin-glass model）, the F-measure of our method is on average improved by 6.34 %, 16.36 % and 12.13 %in the three datasets.