On feature selection clustering by way of multi-objective evolutionary algorithms.

摘要

In this work, we propose a multiobjective evolutionary algorithm, Feature Selection Clustering (FSC), to data clustering with automatic k-determination and capacity to identify good solutions from the pareto front. This NSGA-II based algorithm implements feature selection and generates cluster solutions only with an appropriate subset of features representing the cluster solutions. FSC does not restrict clusters to share the same feature subset but rather allows different clusters be associated with any subset of features, so called subspace clustering. FSC is applied to 5 UCI datasets and the results of FSC are compared to those of X-means and a second clustering model, Random Feature Selection Clustering algorithm (R-FSC), introduced to establish a performance baseline. The algorithms are evaluated based on two objectives (i) distortion value, and (ii) class distribution and our results indicate that FSC produces high quality cluster solutions that meet and on some datasets outperforms X-means algorithm.