Single Classifier-based Multiple Classification Scheme for weak classifiers: An experimental comparison

In this paper, we propose a Single Classifier-based Multiple Classification Scheme (SMCS) that uses only a single classifier to generate multiple classifications for a given test data point. The SMCS does not require the presence of multiple classifiers, and generates diversity through the creation of pseudo test samples. The pseudo test sample generation mechanism allows the SMCS to adapt to dynamic environments without multiple classifier training. Moreover, because of the presence of multiple classifications, classification combination schemes, such as majority voting, can be applied, and so the mechanism may improve the recognition rate in a manner similar to that of Multiple Classifier Systems (MCS). The experimental results confirm the validity of the proposed SMCS as applicable to many classification systems. Even without parameter selection, the average performance of the SMCS is still comparable to that of Bagging or Boosting. Moreover, the SMCS and the traditional MCS scheme are not mutually exclusive, and the SMCS can be applied along with traditional MCS, such as Bagging and Boosting.

► We use only a single classifier to generate multiple classifications for a given test sample.
► The decision boundary optimization is carried out with pseudo test data point generation.
► The method reduces the costs of classifier training and classifier selection for an EoC.
► Diversity can be extracted from datasets directly, without the training of multiple classifiers.
► The extent of diversity can be adjusted by controlling the neighbourhood parameters.