In imbalanced regression tasks, the uneven distribution of continuous target variables poses a significant challenge for predicting rare values, with the core difficulty lying in effectively defining and identifying these rare values. To address this challenge, this paper first proposes a novel rare region identification strategy, named $KK $-means, which combines kernel density estimation and $ K $-means clustering to systematically identify and merge sparse sample points in the target variable space into continuous rare intervals. Furthermore, a rare value bootstrap-based imbalanced regression method (RareBoost) is proposed, based on the identified rare regions. The RareBoost first extracts information from the identified rare intervals through label density ratio weighting and dynamically adjusts sample weights during the bootstrap sampling process to enhance the model’s focus on rare regions. Subsequently, a Stacking meta-learner is employed to integrate these base learners with ‘rare value awareness’, forming a predictive model that balances global efficiency and local accuracy for rare values. Experimental results demonstrate that the RareBoost achieves average improvements of 8.7%, 5.8%, and 13.8% over the best baseline method in terms of ANLL, RMSE, and $ R^{2}$, respectively, confirming its effectiveness.