| Abstract [eng] |
The process of finding features that meet the given constrains out of a large group of features is called feature reduction. The reduction concept can be divided into feature selection and feature extraction techniques. The feature selection approach selects the independent features that provide sufficient information for a satisfactory separation between the different situations we want to discriminate. The physical values of selected features remain unchanged. The redundancy of features might be identified by a feature clustering and selection algorithm or we might remove features with the highest correlation. The algorithm removes similar features. This implies a faster training of consequent classifiers on reduced feature space. The feature extraction method works in opposite. Hereby, the features are projected onto a set of reduced feature space by some transformation function. The features in transformed space are no longer representing the same physical meaning as in original space. The transformation function is an analytical function and the challenge is to find representative and informative transformation for the given feature set. Very well known techniques are: the principal components analysis (PCA) and dimensionality reduction by auto-associative mapping using MLP neural. Four methods for features space reduction were analyzed in this work. All these methods have been used with four publicly available databases and applied to very well known k-nearest neighbor (k-NN) classifier method and the best methods for features space reduction were chosen. In all the tests performed, the feature clustering method was the best, i.e. the least average classification errors was made by feature clustering method. Features space reduction method based on MLP neural achieved the worst (bad) average classification errors. |
