Pearson coefficient matrix for studying the correlation of community detection scores in multi-objective evolutionary algorithm

Abstract

Assessing a community detection algorithm is a difficult task due to the absence of finding a standard definition for objective functions to accurately identify the structure of communities in complex networks. Traditional methods generally consider the detecting of community structure as a single objective issue while its optimization generally leads to restrict the solution to a specific property in the community structure. In the last decade, new community detection models have been developed. These are based on multi-objective formulation for the problem, while ensuring that more than one objective (normally two) can be simultaneously optimized to generate a set of non-dominated solutions. However the issue of which objectives should be co-optimized to enhance the efficiency of the algorithm is still an open area of research. In this paper, first we generate a candidate set of partitions by saving the last population that has been generated using single objective evolutionary algorithm (SOEA) and random partitions based on the true partition for a given complex network. We investigate the features of the structure of communities which found by fifteen existing objectives that have been used in literature for discovering communities. Then, we found the correlation between any two objectives using the pearson coefficient matrix. Extensive experiments on four real networks show that some objective functions have a strong correlation and others either neutral or weak correlations.