New approach to sub-community structure in high-dimensional data

Abstract

For complex systems described as networks, modularity maximization has been emerging as one of community detection methods like PCA and network analysis, due to their intuitive concept and application potential to real systems in spite of the resolution limit. By the traditional methods, however, sub-community structure may not be clearly revealed in many cases. For the complex system of which nodes are expressed in high-dimensional feature vectors, we propose a new procedure using archetypal analysis (AA) and an invented quality function for uncovering the multiscale community structure of the system, visualize the structure with t-SNE and also use machine learning techniques for optimization issues. In this thesis, we study macro- and sub-community structures of various complex systems including generated system, financial system and bioinformatics system with the proposed approach and other traditional methods, and also show that the proposed one can overcome the limit of traditional ones in community detection.