Classification by Cluster Analysis: A New Meta-Learning Based Approach

Combination of multiple classifiers, commonly referred to as an classifier ensemble, has previously demonstrated the ability to improve classification accuracy in many application domains. One popular approach to building such a combination of classifiers is known as stacking and is based on a meta-learning approach. In this work we investigate a modified version of stacking based on cluster analysis. Instances from a validation set are firstly classified by all base classifiers. The classified results are then grouped into a number of clusters. Two instances are considered as being similar if they are correctly/incorrectly classified to the same class by the same group of classifiers. When classifying a new instance, the approach attempts to find the cluster to which it is closest. The method outperformed individual classifiers, classification by a clustering method and the majority voting method.