COMMUNITY DETECTION BY SPECTRAL METHODS UNDER DCSBM AND DCMM

Abstract

This thesis focuses on designing spectral clustering algorithms for solving two problems: community detection based on DCSBM model and mixed membership community detection based on DCMM model. In Chapter 2, we propose six spectral methods Anorm , RSCORE, PCC, NPCC, V and NV for community detection. Theoretic analysis shows that PCC and V yield consistent community detection under mild conditions. Population analysis for Anorm, RSCORE, NPCC and NV show that these four approaches return perfect clustering results for the ideal case under DCSBM. Results for both simulation and eight real world networks show that our methods have satisfactory performances with lower error rates compared with some well-known methods. In Chapter 3, we design seven methods mAnorm, mRSCORE, mRSC, mV, mNV, mPCC and mNPCC for mixed membership community detection problem. Results for both simulation and ten real world networks show that our methods have satisfactory performances.