Evolving Simpler Constructed Features for Clustering Problems with Genetic Programming

Clustering is a widely used unsupervised learning technique. However, as the size and complexity of data increases, the performance of clustering algorithms diminishes, as well as the interpretability of the clustering partition. Genetic programming has been used to perform feature construction on data to increase clustering performance. However, existing work has not focused on encouraging simpler constructed features. In this paper, existing techniques are further developed to include parsimony pressure-a method to encourage evolution towards simpler solutions. With simpler solutions, the constructed features become easier to understand and interpret. The results of experiments using the proposed method show that parsimony pressure is an effective method for producing significantly simpler constructed features without any reduction on the performance of k-means++clustering. Evolved individuals are also analysed to demonstrate the effect of parsimony pressure on interpretability, showing the power of parsimony pressure for avoiding redundancies in individuals, and thus increasing the interpretability.