This paper presents an adaptive genetic clustering method for exploratory mining of feature vector or time series. Our method basically implements the k-medoids algorithm with distance computed based on dynamic time warping for time series of unequal length and Euclidean for feature vectors. Each chromosome encodes objects serving as the k-medoids in binary. Both mutation and crossover rates are adapted during evolution. Six fitness functions are defined based on existing validity indices. An application in clustering grinding signals is first shown to illustrate its use. The method is then further evaluated using benchmark data available in the public domain to test the performance of the fitness functions employed. The results indicate that: (i) the proposed method produces comparable or better clustering accuracies than k-means if there is a priori knowledge about the number of clusters; and (ii) all six validity indices fail to correctly identify the actual number of clusters based on the k-means results for most data sets. A GA-based procedure is proposed to devise a new validity index with two tuning parameters and is shown not only identifying the actual number of clusters for all data sets but also producing better clustering accuracy than k-means and some indices.
