Selection of best clustering features in a wavelet packet library

We develop a new algorithm to select the basis that provides an optimal clustering of non-stationary multidimensional data. This optimality is obtained according to a new cost function that measures the clustering power of a basis vector or feature. Our approach does not require any training sets or any models of the clusters. The basis is searched in a dictionary of wavelet packets. We have conducted experiments with synthetic data. Our method is capable of discovering the structures of the datasets