Performance of symmetric non-negative matrix factorization-based community detector with learning depth variations

Abstract

Community detection is a popular topic in the area of social network analysis. Symmetric non-negative matrix factorization (SNMF) is widely adopted to address this task. This work focuses on connections between the quality of community detection, and the learning strategies for the non-negative and multiplicative update rule (NMU), which is a popular learning scheme for implementing SNMF model. To do so, we adopt α and β-controlled NMU to implement a series of SNMF-based community detectors, and validate their performance on six different social networks generated by industrial applications. Experimental results demonstrate that a) the learning depth of NMU is closely connected with the resultant model's performance in community detection; b) with large α in α-controlled NMU, the resultant SNMF model can achieve the most accurate results in community detection; d) compared with the traditional NMU, both α and β-controlled NMU schemes are able to make an SNMF model achieve significantly higher performance in community detection.