A hybrid high-dimensional feature selection method integrating an improved genetic algorithm and fuzzy neural network

The curse of dimensionality in high-dimensional data remains a critical obstacle for pattern recognition and machine learning under the big data paradigm. To address this challenge, we propose a hybrid feature selection framework (FNN-GA) that integrates an improved Genetic Algorithm (GA) with a Fuzzy Neural Network (FNN), aiming to achieve efficient feature subset exploration and robust classification. The GA component incorporates multi-point crossover, diversity-preserving selection, and dynamically adjusted parameters to enhance global search capability. The FNN component is equipped with stable parameter initialization, dropout-based regularization, and adaptive learning rate scheduling to improve generalization. For fitness evaluation, we design an ensemble strategy that aggregates predictive outcomes from FNN, SVM, and MLP, combined with an adaptive penalty and stability-aware reward mechanism to ensure reliable subset assessment. In addition, a preprocessing strategy leveraging statistical pre-filtering and stratified cross-validation is introduced to further optimize the pipeline. Extensive experiments on six UCI benchmark datasets and a real-world application dataset of steel plates faults show that the proposed FNN-GA consistently outperforms representative baselines, including GA-SVM, PSO-NN, Relief-F, mRMR, DE-SCA, and CAE, in terms of dimensionality reduction and classification accuracy. Statistical significance is validated via the Wilcoxon signed-rank test. Beyond empirical gains, the proposed framework demonstrates strong generalizability and can be readily extended to other domains involving high-dimensional data.