Optimized method of subset generation for ensemble classification

Abstract

Multiclass learning problems involve training classification models to accurately classify data instances into class labels. The number of class labels that are incorporated in a classification model directly affects its training time and accuracy. It has been noted that the existing models, when used on larger datasets having an extensive number of classes, often fail in achieving an accuracy usable for functional, real-world scenarios and tend to overfit on training data. Present research attempts to optimize the performance of classifiers by creating custom classification strategies, and neural network architectures for specific classification tasks and datasets. The efficiency of one such optimization of classifiers, classifier ensembles can further be improved by altering methods of subset generation from the dataset and correspondingly, class prediction. In this regard, we propose a method of constructing ensembles which aims to increase the accuracy of any given classifier while reducing training time by using a two step approach- one step for group formation as a part of constructing ensembles, and another for relative probability calculations which combines the result of ensembles. The proposed method is implemented and experiments are done on the Fashion MNIST (10 class labels), NIH Chest X-Ray + COVID-19 (16 class labels), Kuzushiji-49 (49 class labels), and CIFAR-100 (100 class labels) datasets. Experimental results show the efficacy of the proposed method which has achieved an increase of 6.33%, 9.32%, 8.59%, and 12.27% in accuracies of respective datasets and at the end, results are debated.