Unleashing Deep Reinforcement Learning: A Promising Alternative for Imbalanced Dataset Classification

Class-imbalanced datasets pose a significant challenge for classification tasks in supervised learning, as standard classification algorithms are designed under the assumption that the datasets have balanced class distributions. This can lead to biased models that perform poorly on minority classes. To address this challenge, several approaches have been proposed at both the data level and algorithmic level. In this research study, comparative analysis of Deep Reinforcement Learning (DRL) methods and Oversampling techniques (SMOTE, SVM-SMOTE, ADASYN) along with Deep Learning (DL) algorithms have been performed on two benchmark imbalanced datasets—“Breakhis” and stroke prediction datasets. The experimentation has been carried out by taking different imbalanced ratios of 0.05, 0.1, 0.2, and 0.3 on the Breast Cancer training dataset and 0.05 imbalanced ratio on the Brain Stroke dataset. The comparison of DRL and Oversampling methods has been performed, and the results demonstrate that the DRL model outperformed oversampling techniques on all imbalanced ratios of the “Breakhis” dataset. For the Stroke prediction dataset, Oversampling techniques completely failed in classifying the minority class samples. These findings suggest that the Deep Reinforcement Learning algorithm presents itself as a promising alternative for imbalanced dataset classification in the medical domain and could potentially contribute to improving patient outcomes.