Nearest Centroid Classifier Based on Information Value and Homogeneity

The aim of this paper is to introduce a novel classification algorithm based on distance to class centroids with weighted Euclidean distance metric. Features are weighted by their predictive powers and in-class homogeneities. For predictive power, information value metric is used. For in-class homogeneity different measures are used. The algorithm is memory based but only the centroid information needs to be stored. The experimentations are carried at 45 benchmark datasets and 5 randomly generated datasets. The results are compared against Nearest Centroid, Logistic Regression, K-Nearest Neighbors and Decision Tree algorithms. The parameters of the new algorithm and of these traditional classification algorithms are tuned before comparison. The results are promising and has potential to trigger further research.