Local Feature Selection （LFS） methods partition the sample space into multiple local regions and select the optimal feature subset for each region to reflect local heterogeneous information. However， the existing LFS methods partition local regions around each sample and find the optimal feature subset， resulting in low optimization efficiency and limited applicability. To address this issue， a new evolutionary Bi-level adaptive Local Feature Selection （BiLFS） algorithm was proposed. The LFS problem was formulated as a bi-level optimization problem， with feature subsets and locally optimized regions as the decision variables. At the upper level， Non-dominated Sorting Genetic Algorithm Ⅱ was employed to find the optimal feature subsets for the selected local regions， with region purity and selected feature ratio as the objective functions. At the lower level， based on the upper-level solution， local region clustering analysis was used to obtain center samples within each region， followed by local region fusion to eliminate unnecessary regions and update the population of necessary regions. Experimental results on 11 UCI datasets demonstrate that BiLFS achieves an average classification accuracy up to 98.48%， and an average computation time down to 9.51% compare to those of non-adaptive LFS methods based on evolutionary algorithms， significantly improving computational efficiency to the level of linear programming-based LFS methods. Visual analysis of the locally optimized regions selected by the BiLFS algorithm during the iteration process indicates the stability and reliability of selecting necessary local regions.