1. Subconcept perturbation-based classifier for within-class multimodal data.
- Author
-
Cavalcanti, George D. C., Soares, Rodolfo J. O., and Araújo, Edson L.
- Subjects
- *
CLASSIFICATION algorithms - Abstract
In classification, it is generally assumed that data from one class consist of one pure compact data cluster. However, in many cases, this cluster might consist of multiple subclusters, in other words, within-class multimodality. In such a scenario, it may be difficult or even impossible for a single classifier to find a suitable model using limited data. So, training a model using smaller chunks of data is an alternative that helps avoid complex models and reduces the task's complexity. This paper proposes the subconcept Perturbation-based Classifier (sPerC) that finds the best clusters per class using cluster validation measures, and one meta-classifier is trained per subcluster. This way, each class is represented by a set of meta-classifiers instead of one classifier. Such a design diminishes the complexity of the task, and using a divide-to-conquer strategy favors the precision of each meta-classifier. Through a set of comprehensive experiments on 30 datasets, the sPerC results compared favorably to other classifiers in multi-class classification tasks, showing that creating specialized classifiers per class in different regions of the feature space can be advantageous. [ABSTRACT FROM AUTHOR]
- Published
- 2024
- Full Text
- View/download PDF