Exploring Mixed Membership Stochastic Block Models via Non-negative Matrix Factorization

Chengbin Peng, Ka Chun Wong

Research output: Chapter in Book/Report/Conference proceedingConference contribution


Many real-world phenomena can be modeled by networks in which entities and connections are represented by nodes and edges respectively. When certain nodes are highly connected with each other, those nodes forms a cluster, which is called community in our context. It is usually assumed that each node belongs to one community only, but evidences in biology and social networks reveal that the communities often overlap with each other. In other words, one node can probably belong to multiple communities. In light of that, mixed membership stochastic block models (MMB) have been developed to model those networks with overlapping communities. Such a model contains three matrices: two incidence matrices indicating in and out connections and one probability matrix. When the probability of connections for nodes between communities are significantly small, the parameter inference problem to this model can be solved by a constrained non-negative matrix factorization (NMF) algorithm. In this paper, we explore the connection between the two models and propose an algorithm based on NMF to infer the parameters of MMB. The proposed algorithms can detect overlapping communities regardless of knowing or not the number of communities. Experiments show that our algorithm can achieve a better community detection performance than the traditional NMF algorithm. © 2014 IEEE.
Original languageEnglish (US)
Title of host publication2014 IEEE International Conference on Data Mining Workshop
PublisherInstitute of Electrical and Electronics Engineers (IEEE)
Number of pages6
ISBN (Print)9781479942749
StatePublished - Dec 2014

Bibliographical note

KAUST Repository Item: Exported on 2020-10-01


Dive into the research topics of 'Exploring Mixed Membership Stochastic Block Models via Non-negative Matrix Factorization'. Together they form a unique fingerprint.

Cite this