A SOM-based dimensionality reduction method for KNN classifiers

The self-organizing-feature-maps (SOM) algorithm is a typical dimensionality reduction technique. The SOM algorithm adopts neighborhood learning to form a topological ordering among data points. In other words, self-organizing feature maps highly preserve topological relationships in the lower-dimensional space. Using SOM as a feature extraction method for the k nearest neighbor classifier is appropriate, since we always choose k ordered samples in the classification phase. This paper uses self-feature-maps to represent original data sets in a two-dimensional feature space in the learning phase to reduce classification time of the k nearest neighbor classifier. Since the self-organizing feature maps algorithm preserves distance and proximity relationships, our proposed method does not compromise k nearest neighbor classification accuracy, but obtains better k NN classification accuracy in lesser time. This work proposes a weighted-self-organizing feature maps (WSOM) method using a weighted distance of finding the winning neuron step. Experiments with artificial datasets and real datasets verify the proposed method performance. Experimental results show that our proposed algorithm performs the best and is most efficient at the classification phase.