Recursive Harmony Search Based Classifier Ensemble Reduction

In recent times classifier ensembles have become a mainstay in data mining and machine learning. The combination of several classifiers generally results in better performance and accuracy as compared to a single classifier. There are many different methods and techniques for constructing ensembles. Most of the time however, when these ensemble classifiers are constructed, the data used in the construction of ensemble classifiers becomes redundant. This redundant data results in a loss of accuracy and an increase in memory and system overhead. Therefore by removing this redundant data we can reduce the memory and system overhead as well as obtain an increase in accuracy. The redundant data can be eliminated by using a technique called feature selection. Feature selection is used to select the most relevant features while performing any task. There are many different feature selection algorithms such as memetic algorithms, sub-modular feature selection, etc. The feature selection technique can be used to choose the relevant data and eliminate the redundant data. The way to eliminate redundant data in ensemble classifiers is to perform classifier ensemble reduction. This paper discusses using feature selection and in particular employing recursive harmony search to perform classifier ensemble reduction via feature selection. The final ensemble classifier will be a reduced set of the original ensemble classifier, while maintaining diversity and accuracy of the original one.