A study of automatic clustering based on evolutionary many-objective optimization

Automatic clustering problems, which need to detect the appropriate clustering solution without a pre-defined number of clusters, still remain challenging in unsupervised learning. In many related works, cluster validity indices (CVIs) play an important role to evaluate the goodness of partitioning of data sets. However, there is no CVI that is likely to ensure reliable results for different structures of data. In this paper, we present a study of evolutionary many-objective optimization (EMaO) based automatic clustering, in contrast to the weighted sum validity function defined in literature, several validity functions (more than 3) are considered to be optimized simultaneously here. Since the research of EMaO is still in its fancy, we take four state-of-the-art EMaO algorithms into consideration as the underlying optimization tool. To be more applicable and efficient for clustering problems, the encoding scheme and genetic operators are redesigned. Experiments show that, for the purpose of this study, it is promising to address automatic clustering problems based on a suitable EMaO approach.