High-dimensional imbalanced datasets, characterized by elevated dimensionality and significant class imbalance, which presents substantial challenges to conventional classification algorithms. This imbalance induces classifier bias towards majority classes, while redundant features exacerbate performance degradation. To mitigate these issues, first, we propose a multi-objective feature extraction framework incorporating class separability and feature generalization performance as optimization objectives, explicitly accounting for class imbalance within the objective formulation. Then, we propose a bilevel optimization-based decision tree algorithm in which non-leaf nodes implement bilevel optimized classifiers. The upper level performs feature subspace exploration, while the lower level derives the optimal class-separating hyperplane for the identified subspace. Comparative evaluation on benchmark datasets demonstrates that the proposed methodology achieves statistically significant superiority over state-of-the-art approaches in both F-score and G-mean metrics, thereby validating its efficacy.