Feature extraction by neural network nonlinear mapping for pattern classification

Department of Electrical and Computer Engineering Ben-Gurion University of the Negev Beer-Sheva 84105, Israel Abstract Feature extraction has been always mutually studied for exploratory data projection and for classification. Feature extraction for exploratory data projection aims for data visualization by a projection of a high-dimensional space onto two or three-dimensional space, while feature extraction for classification generally requires more than two or three features. Therefore, feature extraction paradigms for exploratory data projection are not commonly employed for classification and vice versa. We study extraction of more than three features, using neural network (NN) implementation of Sammon’s nonlinear mapping to be applied for classification. Comparative classification experiments reveal that Sammon’s method, which is primarily an exploratory data projection technique, has a remarkable classification capability. The classification performance of (the unsupervized) Sammon’s mapping is highly comparable with the performance of the principal component analysis (PCA) based feature extractor and is slightly inferior to the performance of the (supervized) multilayer perceptron (MLP) feature extractor. The paper thoroughly investigates a random and a non-random initializations of Sammon’s mapping. Only one experiment of Sammon’s mapping is required when the eigenvectors corresponding to the largest eigenvalues of the sample covariance matrix are used to initialize the projection. This approach tremendously reduces the computational load and substantially raises the classification performance of Sammon’s mapping using only very few eigenvectors.