Applications of Autonomous Learning Multi Model System to Multiclass Imbalanced Datasets

André Seabra, Rodrigo M. M. Ventura, Rui Jorge Almeida, Susana M. Vieira, João M. C. Sousa

Abstract: Imbalanced datasets are datasets with a class imbalance problem, where the number of instances of one class is much higher than the number of instances of the other classes. This causes issues in classifiers, which tend to be biased towards the largest classes. The imbalanced dataset problems cause extra challenges when dealing with more than two-classes, due to possible skewness and other distribution artifacts in the different classes. In this paper, we study the zero-order autonomous learning multi-model (ALMMo-0) classifier for multiclass imbalanced datasets. The ALMMo-0 is self-organising nonparametric fuzzy rule-based model proposed for classification problems, without any explicit mechanism for imbalanced datasets. The ALMMo-0 has shown good performance in multiclass problems but also in binary imbalanced data, but it was never systematically studied for the problem of multiclass imbalanced datasets. Furthermore, we propose an extension of the ALMMo-0 using class decomposition, a general way to deal with class imbalance problems and multiclass problems, in an ensemble form and in a sequential manner. We compare the performance of the ALMMo-0 classifier with other well studied classification methods, also without any explicit mechanisms for class imbalance. Our results show that the proposed methods of class decomposition improve the results of the ALMMo-0 classifier for imbalanced datasets, and all experiments help to provide a better understanding of how the algorithm learns in these datasets.