Oversampling method based on angle and direction clustering

Abstract

The invention discloses an oversampling method based on angle and direction clustering, and the method comprises the steps: obtaining an unbalanced data set, carrying out the clustering of the unbalanced data set through employing a clustering algorithm, generating a clustering label, an angle variance, and a sorting neighbor set for each sample in the unbalanced data set, and carrying out filtering processing on each sample of which the clustering label is noise so as to obtain a filtered sample; and calculating a first oversampling weight, a second oversampling weight and an optimal interpolation neighbor set of each minority class sample in the unbalanced data set according to the clustering label, the angle variance and the sorting neighbor set of each sample in the unbalanced data set, and calculating the oversampling weight of each cluster in the unbalanced data set and the number of new samples needing to be synthesized by each cluster according to the first oversampling weights of all minority class samples. The technical problem that the importance of boundary samples in classification is ignored in an existing oversampling method can be solved.

Description

technical field [0001] The invention belongs to the technical field of data mining, and more particularly, relates to an oversampling method based on angle and direction clustering. Background technique [0002] The development of machine learning and deep learning provides powerful support for classification and prediction. In classification problems, as a kind of supervised learning, it is necessary to provide a data set for the classification model for model training, however, the data set is unbalanced in most cases. The samples with a higher proportion of labels in the dataset are called majority class samples, and the samples with lower label proportions are called minority class samples. Imbalance means that the number of majority class samples in the dataset is often far more than the minority class samples. Minority class samples have a small amount of data, which makes the classifier unable to learn minority class samples effectively, and then it is difficult to m...

AI Technical Summary

Problems solved by technology

[0005] In view of the above defects or improvement needs of the prior art, the present invention provides an oversampling method based on angle and direction clustering, the purpose of which is to solve the problem of ignoring the importance of boundary samples in classification in the existing oversampling method Technical problems, and because the relationship distribution between samples is ignored, the synthetic samples are easy to fall between the majority class samples, which leads to the technical problems of low quality synthetic samples, and because the distance and density measurement methods will be in high-dimensional space will worsen, making the method less effective in high-dimensional imbalanced datasets.

Images