Adaptive Q-Learning for imbalanced data: A framework with dynamic autoencoders for complexity reduction
Abstract
Imbalanced classification with high data complexity, characterised by poor class separability and feature overlap, remains a critical challenge in machine learning, as conventional models often bias towards majority classes. This paper proposes ε-DQN-B-A-D, a novel two-stage framework that incorporates feature transformation with adaptive reinforcement learning. In the first stage, an autoencoder guided by a dynamically weighted loss function balances reconstruction error with Fisher’s Discriminant Ratio to enhance feature representation and preserve minority class information. In the second stage, a boosted Deep Q-Learning agent (ε-DQN-B) exploits an XGBoost model to adaptively regulate the ε- greedy exploration rate, improving sensitivity to minority classes. Evaluations on 24 benchmark datasets demonstrate that ε-DQN-B-A-D achieves superior performance, significantly outperforming baseline and state-of-the-art methods in G-mean. Statistical tests (Friedman and Nemenyi) confirm its robustness, with a ranking of 2.2708 compared to 3.0625 and 3.9375 for competitive baselines. These findings establish ε-DQN-B-A-D as an effective solution for complex, imbalanced classification tasks.