Density Weighted Region Growing Method for Imbalanced Data SVM Classification in Under-sampling Approaches

摘要

A density-weighted under-sampling method for SVM on imbalanced data is proposed. To reduce the size of majority group in training set, the region growing clustering method is employed to partition data into several clusters and the cluster centers are considered as the representatives of majority group. To initialize seeds of region growing, the density of each point in majority group is calculated first. The seeds then are randomly picked up and the probability of one point to be seed is in proportion with its corresponding density. The under-sampled training set is built by the minority group and the representatives of majority group. Experimental results on toy data and nature data show the priority of the proposed method comparing with Randomly under-sampling method and CNN.