Improving k-means Clustering with Genetic Programming for Feature Construction

k-means is one of the most commonly used clustering algorithms in data mining. Despite this, it has a number of fundamental limitations which prevent it from performing effectively on large or otherwise difficult datasets. A common technique to improve performance of data mining algorithms is feature construction, a technique which combines features together to produce more powerful constructed features that can improve the performance of a given algorithm. Genetic Programming (GP) has been used for feature construction very successfully, due to its program-like structure. This paper proposes two representations for using GP to perform feature construction to improve the performance of k-means, using a wrapper approach. Our results show significant improvements in performance compared to k-means using all original features across six difficult datasets.