Dynamic selection and combination of one-class classifiers for multi-class classification

A natural solution to tackle multi-class problems is employing multi-class classifiers. However, in specific situations, such as imbalanced data or high number of classes, it is more effective to decompose the multi-class problem into several and easier to solve problems. One-class decomposition is an alternative, where one-class classifiers (OCCs) are trained for each class separately. However, fitting the data optimally is a challenge for OCCs, especially when it presents a complex intra-class distribution. The literature shows that multiple classifier systems are inherently robust in such cases. Thus, the adoption of multiple OCCs for each class can lead to an improvement for one-class decomposition. With that in mind, in this work we introduce the method called One-class Classifier Dynamic Ensemble Selection for Multi-class problems (MODES, for short), which provides competent classifiers for each region of the feature space by decomposing the original multi-class problem into multiple one-class problems. So, each class is segmented using a set of cluster validity indices, and an OCC is trained for each cluster. The rationale is to reduce the complexity of the classification task by defining a region of the feature space where the classifier is supposed to be an expert. The classification of a test example is performed by dynamically selecting an ensemble of competent OCCs and the final decision is given by the reconstruction of the original multi-class problem. Experiments carried out with 25 databases, 4 OCC models, and 3 aggregation methods showed that the proposed architecture outperforms the literature. When compared with the state-of-the-art, MODES obtained better results, especially for databases with complex decision regions. (C) 2021 Elsevier B.V. All rights reserved.

Automated summary

This work introduces a new method called MODES, which decomposes the original multi-class problem into multiple one-class problems to provide competent classifiers for each region of the feature space. Experimental results show that this method outperforms the literature, especially for databases with complex decision regions.