METHOD FOR CLASSIFYING HIGH-DIMENSIONAL IMBALANCED DATA BASED ON SVM

摘要

A method for classifying high-dimensional imbalanced data based on SVM, comprising two parts. The first part is feature selection, involving: using an SVM-BRFE algorithm to resample a boundary to look for an optimal feature weight so as to carry out feature importance measurement, feature selection and training set update, and repeating the process. Finally, the feature most conductive to enhancing an F1 value is retained, and other features are removed, so that a subsequent training process is carried out in a situation with feature redundancy and irrelevant feature combination as less as possible and dimension as low as possible, thereby reducing the influence of a high dimension problem on an imbalance problem and the constraint over an SMOTE oversampling algorithm. The second part is data sampling, involving: using an improved SMOTE algorithm, i.e. a PBKS algorithm; considering to use minority classes in boundaries automatically partitioned by SVM as distance constraints in D^Hx_ij of a Hilbert space so as to replace an original constraint; and using a grid method to look for the approximate preimage. The method can stably and effectively complete the task of classifying high-dimensional unbalanced data, and can achieve a considerable effect.