Combining Wavelets and Computational Intelligence Methods with Applications on Multi-class Classification Datasets

In this paper, we propose a novel algorithm for wavelet feature extraction as input to a supervised Multi-Class Classifier to improve classification performance. In particular, to select the best wavelets coefficient features, we first compute the energy-based variance distribution from wavelets coefficients at different subbands as well as the entropy-based fuzzy measures associated with the training instances. Once we get these entropy-based fuzzy measures associated with the different subsets of wavelets subbands, we apply the Möbius Transform to these entropy-based fuzzy measures to extract the Multivariate Mutual Information associated with the different subsets of wavelets subbands. The goal of these measures is twofold: assign weights (based on the wavelets information content) to all subsets of wavelets subbands and extract the independent (in terms of the multivariate mutual information) subsets of wavelets subbands. In our case, the optimal subsets of wavelets subbands as wavelets features vectors to train a Bayesian Network Model are those which provide a multivariate mutual information equal to zero. Experimental results with the multi-class SRBCT cancer dataset, show that our proposed approach achieves lower classification error in comparison with other methods proposed in the literature. Keywords— fuzzy measures, mutual information, wavelets.