Imbalanced Regression Pipeline Recommendation

Abstract

Imbalanced problems are common in various real-world scenarios and present significant chal- lenges, especially for regression tasks due to the rarity of certain continuous target values. While these issues have been extensively explored in classification tasks, they also affect re- gression, complicating model performance. This work presents an extensive experimental study involving various balancing strategies and learning models, introduces a taxonomy for imbal- anced regression approaches based on regression models, learning process modification, and evaluation metrics, and highlights new insights into the advantages of different strategies. From this study, it became evident that the choice of resampling method depends on the problem, learning models, and metrics, making it difficult to select an appropriate resam- pling strategy and learning model. As a result, it is necessary to test the majority of existing combinations. Based on these findings, this work proposes the Meta-learning for Imbalanced Regression (Meta-IR) framework to address these challenges. Meta-IR recommends optimal pipelines consisting of resampling strategies and learning models for imbalanced regression tasks. Two formulations are proposed: Independent, which separately recommends learning algorithms and resampling strategies, and Chained, which models their interdependencies se- quentially. The Chained approach demonstrated superior performance, suggesting a significant relationship between learning algorithms and resampling strategies. Compared with AutoML models and baseline configurations, Meta-IR outperformed all, offering a more effective solu- tion for imbalanced regression and indicating directions for future research.