A novel similarity classifier with multiple ideal vectors based on k-means clustering

摘要

In the literature, researchers and practitioners can find a manifold of algorithms to perform a classification task. The similarity classifier is one of the more recently suggested classification algorithms. In this paper, we suggest a novel similarity classifier with multiple ideal vectors per class that are generated with k-means clustering in combination with the jump method. Two approaches for pre-processing, via simple standardization and via principal component analysis in combination with the MAP test and Parallel Analysis, are presented. On the artificial data sets, the novel classifier with standardization and with transformation power Y = 1 for the jump method results in significantly higher mean classification accuracies than the standard classifier. The results of the artificial data sets demonstrate that in contrast to the standard similarity classifier, the novel approach has the ability to cope with more complex data structures. For the real-world credit data sets, the novel similarity classifier with standardization and Y = 1 achieves competitive results or even outperforms the k-nearest neighbour classifier, the Naive Bayes algorithm, decision trees, random forests and the standard similarity classifier.