Feature Selection Research Based On Maximum Relevance Minimum Redundancy

Feature selection is a key problem in pattern recognition field which selects optimal feature subset from the original characteristics. For example, in the field of bioinformatics, a more efficient feature selection algorithm is especially important in small samples of high-dimensional data for gene expression and protein mass spectrometry data.Feature selection is necessary in designing a good classifier performance. For example, the computational complexity and training time of support vector machine classifiers is a non-linear change with the number of training samples and input space dimension. Therefore, preprocessing on the training set is an important way to improve the performance of SVM. Choosing reasonable and effective features and reducing dimension appropriately can eliminate redundancy and accelerate the running speed to improve the classification efficiency, on the other hand, it can reduce the complexity of the classifier and error rate.In order to solve the problems of algorithm complexity and determine the number of best characteristics in feature selection process, the paper proposes an improvement wrapper type feature selection algorithms. It can achieve the best characteristic quantities measure choice based on guidelines of feature subset.The paper combined with feature relevance and redundancy proposes an improvement wrapper type feature selection algorithms based on the maximum relevance minimum redundancy feature selection algorithm.The algorithm considers the different effect of relevance and redundancy degree in feature selection, and pulls in a weighting factor to balance relevance and redundancy characteristics. The paper uses the UCI data sets to verify the validity of algorithm, the results show that the algorithm can effectively remove unrelated redundant features, and measures potential redundancy feature space effectively, while also reduces dimension and improves the classification accuracy.