Search

In order to improve the classification performance of a single classification model, Multiple Classifier Systems (MCS) are used. One of the most common techniques utilizing multiple decision trees is the random forest, where diversity between base classifiers is obtained by bagging the training dataset. In this paper, we propose the algorithm that uses horizontal partitioning the learning set and uses decision trees as base models to obtain decision regions. In the proposed approach feature space is divided into disjoint subspace. Additionally, the location of the subspace centroids, as well as the size and location of decision regions, are used in order to determine the weights needed in the last process of creating MCS, i.e. in the integration phase. The proposed algorithm was evaluated employing multiple open-source benchmarking datasets, compared using accuracy and Matthews correlation coefficient performance measures with two existing MCS methods - random forest and majority voting. The statistical analysis confirms an improvement in recognition compared to the random forest. In addition, we proved that for infinitely dense space division proposed algorithm is equivalent to majority voting.

A vital aspect of the Multiple Classifier Systems construction process is the base model integration. For example, the Random Forest approach used the majority voting rule to fuse the base classifiers obtained by bagging the training dataset. In this paper we propose the algorithm that uses partitioning the feature space whose split is determined by the decision rules of each decision tree node which is the base classification model. After dividing the feature space, the centroid of each new subspace is determined. This centroids are used in order to determine the weights needed in the integration phase based on the weighted majority voting rule. The proposal was compared with other Multiple Classifier Systems approaches. The experiments regarding multiple open-source benchmarking datasets demonstrate the effectiveness of our method. To discuss the results of our experiments, we use micro and macro-average classification performance measures.