Reproduce “Deep Reinforcement Learning for Multi-class Imbalanced Training” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jenny Yang, Rasheed El-Bouri, Odhran O'Donoghue, Alexander S. Lachapelle, Andrew A. S. Soltan, David A. Clifton
A deep reinforcement learning framework, Q-imb, utilizing Double Dueling Deep Q-Networks and a custom reward function, effectively addresses extreme multi-class imbalanced classification. It yields superior balanced classification performance, especially for minority classes, and demonstrates strong generalizability across clinical diagnostic tasks without requiring post-hoc threshold adjustments.
Key insights
- Deep Reinforcement Learning, particularly with Dueling architectures and tailored reward functions, can effectively learn augmented representations that improve classification performance on extremely imbalanced and multi-class datasets.
- The dueling network architecture significantly enhances DRL's ability to learn robust policies for classification by decoupling state and action value estimation, outperforming non-dueling DRL methods.
- A carefully designed episode termination rule, which penalizes minority class misclassifications, provides a strong learning signal that prioritizes accurate prediction of underrepresented classes without requiring post-hoc threshold adjustment.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.