An integrated non-parametric sequential approach to multi-class pattern classification
Belur V. Dasarathy · International Journal of Systems Science · 1973
A now method of pattern classification, evolved by integrating in a sequential mode a non-parametric feature selection criterion with an explicit learning scheme, is presented. This feature selection criterion is based on the well-known concept of inter-class and intra-class Euclidean distances as a measure of the separability of the pattern classes in a given feature space. An ‘ effective figure of merit’ is denned and the feature subset in which this figure of merit attains the maximum value is construed as the best feature subset. Usually in most of the existing techniques, a single feature subset is chosen as the hest for the multi-class problem as a whole. A distinctive departure from this practice has been made here in that an individual best feature subset is determined for each of the pattern classes. The values of the effective figure of merit for the best feature subsets of the different pattern classes are sorted to determine the best separable pattern class. The learning scheme developed here is employed in conjunction with the feature selection scheme to train the weight vector of the beat separable class. This learning is explicit in that the weight vector is moved in the appropriate feature subspace to search for the position of minimum classification error amidst the training sample clusters with the errors being monitored explicitly throughout the learning phase. This weight vector is then used to classify the training set and those samples (from the remaining classes) which are classified as unknown form the reconstituted training sample set. This procedure of selecting the best feature subset of the best separable pattern class and training the weight vector of that class is continued sequentially until alt the pattern classes have been considered.