Clustering: Evolutionary Approaches

This thesis is concerned with exploratory data analysis by means of Evolutionary Computation techniques. The central problem addressed is cluster analysis. The main challenges arisen from the unsupervised nature of this problem are investigated. Clustering is a problem lacking a formal general-accepted objective. This justifies the multitude of approaches proposed in literature. A review of the main clustering algorithms and clustering objectives is made. A new approach that takes into account both global and local distribution in data is proposed with the aim of combining the strengths of two different clustering paradigms: centroid-based approaches and density-based approaches. The use of distance metrics in cluster analysis and their impact on the solution space are discussed. The field of metric learning is reviewed. Special emphasis is placed on feature selection methods that aim at extracting a lower-dimensional manifold from data, manifold that maximizes the clustering tendency in data. A wrapper scenario based on multi-modal search evolutionary algorithms is investigated in order to identify feature subsets relevant for the clustering task. A new clustering criterion is formulated able to offer a ranking of partitions derived in feature subspaces of different cardinalities. Particular clustering problems are approached with Evolutionary Computation techniques. Community detection in social networks based on local trust metrics raise a new challenge to clustering analysis: the underlying feature space can not be transformed straightforward into a metric space. Graph clustering is formulated as a multi-objective problem in order to address important applications in VLSI design.