Improvement Of Feature Selection Method Based On Genetic Algorithm

With the increasing of the complexity of researched objects, the dimension of features involved becomes higher and higher, and the redundancy of features increases correspondly. It is hotspot to use feature selection algorithms to ruduce the redundancy and improve the efficiency and accuracy of pattern recognization system now. This dissertation discusses the current relevant research situation, basic theory and methods of feature selection, and emphasizes three aspects of feature selection algorithnm: selection mode, search strategy and evaluation criterion. The major contributions of the dissertation are as follows:1) Parallel multi-criteria feature selection algorithm based on improved genetic algorithm is proposed. First of all, according to the low accuracy and prematurity of basic genetic algorithm used for feature selection, some improvement are done: by using adaptive crossover and mutation operators, and introducing the competition between link-like agents, the convergence speed and optimal solution have been improved. Then, the improved genetic algorithm is used for feature selection. This paper proposed a parallel mode multi-criteria feature selection algorithm to improve the performance of feature selection algorithm with single criterion, the concrete approaches are: select the optimal feature subset based on every single criterion, then choose final feature subset throuth a proper evaluation mechanism. The experimental results show that parallel multi-criteria feature selection algorithm is better than single criterion one, the former can effectively remove redundant features, reduce feature dimension and improve classification accuracy.2) Poll mode and multi-criteria feature selection algorithm based on improved genetic algorithm is proposed. Differently from parallel mode, this poll mode uses every single criterion in sequence, so that the former criterion can be reinforced and amended by latter ones. The experimental results show that poll mode and multi-criteria feature selection algorithm performs better than single criterion algorithms on accuracy.3) Proportional hybrid mode feature selection algorithm is proposed. The author combines filter and wrapper mode and proposes a proportional hybrid mode feature selection algorithm: first, the filter mode feature selection based on genetic algorithm is adopted based on some evaluation criterion; then the individuals with higher fitness value are picked out in proportion to be evaluated again with wrapper mode, This process is repeated for many times until satisfactory feature subset is found. The experimental results show that the accuracy of proportional hybrid mode feature selection algorithm is almost as good as wrapper mode one, and is much better than filter, while the time cost is far less than wrapper mode one. Besides, the stability of hybrid mode is better than both wrapper mode and filter mode.