Research On Feature Selection Methods

Feature selection is one of the key techniques in target recognition. There are mainly three methods of feature selection as follows: embedded methods, filter methods and wrapper methods. The three methods are different according to their combination with classifier. Filter methods and wrapper methods are mainly researched here.Firstly, the feature selection frame proposed by Dash and Liu is quoted, which indicates that feature selection methods include four steps: a generation procedure to generate the next candidate subset, an evaluation function to evaluate the subset under examination, a stopping criterion to decide when to stop and a validation procedure to check whether the subset is valid. Search strategies and evaluation functions are summarized based on the frame in the thesis.Filter methods adopt evaluation functions independent of the classifier. Two filter methods are researched in this paper. First, a combined method is introduced. The feature subset produced by ReliefF algorithm is efficient for classifying, but the features selected may be redundant. Mitra proposes an algorithm based on the maximal information compression index, which has a good performance in removing redundant features. But the selected features may not be the best for the classifier. The combined algorithm performs well in both aspects. The other filter method is improved HFR algorithm. To overcome the shortcoming of HFR, that is CR value is used for evaluating the importance of candidate features, which is related with SDM, we take SGF value as importance evaluation of candidate features instead of CR value.The evaluation functions of wrapper methods are dependent on the classifier. In this paper, two new wrapper methods are proposed. The first one is based on recognition result matrix. Enlightened by the discrimination matrix of Rough set, the concept of recognition result matrix is proposed. Based on this matrix, features are selected using the idea of making MDL. The results of the experiments demonstrate the algorithm can improve the speed and accuracy of the classifier. The second algorithm is based on the complementary coefficient. Complementary coefficient is proposed at the aspect of the interplay of classification ability of features. Based on complementary coefficient, the algorithm selects features from candidate features which have heavier weights. The algorithm has a good performance in removing ineffective features and redundant features. Otherwise, compared with general wrapper methods, the new algorithm has lower time complexity.