A genetic feature weighting scheme for pattern recognition

This paper proposes a new pattern recognition scheme, combining a new adaptive feature weighting and modified k-Nearest Neighbor (k-NN) rule. The proposed feature weighting method named adaptive-3FW. It uses three non-uniform weight levels (zero weight, middle weight and full weight) to weight each feature. The middle weight value is determined using genetic algorithms (GAs). The proposed adaptive-3FW overcomes overfitting issues and achieves high recognition performance. Novel GA operators tailored for this formulation are introduced to implement the proposed scheme. Further, a modified k-NN is proposed which uses a class-dependent feature weighting strategy. Whilst the conventional pattern recognition systems use the same set of feature weights for all classes, the proposed algorithm uses different sets of feature weights for different classes. Experiments were performed with the UCI repository for machine learning databases and the unconstrained handwritten numeral database of Concordia University in Canada to show the performance of the proposed method.