Feature selection (FS) is an effective data pre-processing method that solves the dimensionality disaster caused by data redundancy by selecting a set of features with high relevance and low redundancy in high-dimensional data. Many computational methods have been applied to solve the FS problem, among which the teaching and learning-based optimization algorithm (TLBO) feature selection model has received increasing attention from scholars due to its efficient global search capability. However, with the increasing size of data, the limitations of these algorithms, such as model instability, low model accuracy and poor local search ability, have gradually put the research of the algorithms into difficulties. To address these problems, this paper proposes a hybrid evolutionary Wrapper algorithm model (Teaching and Learning-Based Optimization- Local Search algorithm,TLBOLS) that integrates teaching-learning optimization algorithms with local search methods. Firstly, the algorithm converts the real-type coding to binary coding in the initialization phase, then introduces the worst individual restart mechanism in the teaching phase, and proposes a binary teaching-learning feature selection algorithm for the evolutionary class process using different values of TF values for the two identities of learners and pedagogues (Binary Teaching and Learning-Based Optimization- Local Search algorithm, BTLBOLS). Subsequently, a local search method combining multiple operations and variable neighborhood search is proposed to gradually enhance the perturbation strength and improve the individual quality of the whole population. To optimize the feature selection results, BTLBOLS utilizes a comprehensive evaluation metric as an objective function to guide the overall evolutionary process. Forty-five high-dimensional cancer gene expression datasets are selected for testing and compared with ten feature selection algorithms, and the experimental results show that compared to other algorithms, the BTLBOLS has certain advantages in terms of classification accuracy and number of features, which effectively improves the algorithm classification performance. © 2023 Chinese Institute of Electronics. All rights reserved.
