Let Multi-classification Help Deep Imbalanced Regression

Dekun Lin, Tailai Peng, Rui Chen, Xinran Xie, Zhe Cui

Abstract: Regression is a fundamental task in machine learning. Traditional methods predict target values directly, which limits the robustness of models. More recent methods draw inspiration from human intuition, predicting intervals rather than direct values. However, completely transforming regression tasks and relying solely on a single classifier restrict the contribution of classification to regression. Moreover, data imbalance is widespread in real-world regression tasks, with many labels having insufficient samples. This leads to a lack of gradients for those samples and a significant decline in the performance of deep neural networks. Another challenge lies in generating and amplifying gradients for few-shot samples in a reasonable and effective manner. To address the problems mentioned above, we propose incorporating classification as an auxiliary task to provide additional gradients for few-shot samples. Leveraging multi-grained classifiers aids in achieving more robust features and more accurate regression. Extensive experiments conducted on three benchmark datasets demonstrate that results from our method surpass those of prior state-of-the-art approaches, evidencing the effectiveness of our strategy.