Automated algorithm composition of unsupervised image clustering algorithms

Abstract

Unsupervised learning algorithms are popular as they do not require annotated data. However as per the no-free lunch theorem, the best algorithm to use is not the same for all datasets. This study is the first to automate the composition of an unsupervised image clustering algorithm. This work uses two different techniques to perform automated algorithm composition. The first technique is a genetic algorithm (GA) and the second is a genetic algorithm hyperheuristic (GAHH). A comparison of the two techniques shows that the GA outperforms the GAHH. The GA designs unsupervised clustering algorithms that result in state of the art performance for the Oral lesion, Celebrity faces and COVID-19 datasets.