Abstract:
A variable-length particle swarm optimization (VLPSO) shows good performance for feature selection on large data sets. However, its completely random particle initialization will result in certain blindness in the initial stage. Meanwhile, the single updating mechanism of VLPSO and the information isolation among subpopulations also affect the classification performance. In order to cope with the defect of VLPSO, this paper proposes a co-evolutionary feature selection method based on variable-length particle and multi-behavior interaction(M-CVLPSO). Firstly, to improve the blindness caused by random initialization, the multidirectional initialization strategy in continuous space is adopted to shorten the distance between the initial solution and the optimal solution from the perspective of expectation. Secondly, particles are divided into leaders, followers, and weeders according to fitness, and then multiple updating strategies are adopted in the process of iteration to balance the diversity and convergence of dynamic algorithms. At the same time, the dimension reduction index is integrated into the fitness function to further enhance the performance of the algorithm on some datasets. The convergence of the proposed algorithm is proved theoretically. Finally, the experimental analysis is carried out on the classification accuracy, dimension reduction and calculation time based on 11 large-scale feature selection data sets, which show that the proposed model has better comprehensive performance than the four comparison algorithms.