Genetic Programming for Evolving Similarity Functions Tailored to Clustering Algorithms

Clustering is the process of grouping related instances of unlabelled data into distinct subsets called clusters. While there are many different clustering methods available, almost all of them use simple distance-based (dis)similarity functions such as Euclidean Distance. However, these and most other predefined dissimilarity functions can be rather inflexible by considering each feature equally and not properly capturing feature interactions in the data. Genetic Programming is an evolutionary computation approach that evolves programs in an iterative process that naturally lends itself to the evolution of functions. This paper introduces a novel framework to automatically evolve dissimilarity measures for a provided clustering dataset and algorithm. The results show that the evolved functions create clusters exhibiting high measures of cluster quality.