Hybrid sampling method based on boundary samples

Abstract

The invention relates to a hybrid sampling method based on boundary samples, which comprises the following steps: firstly, dividing majority class samples into boundary samples and internal samples through a boundary point detection algorithm based on a variable coefficient, and reversely determining minority class boundary samples according to the majority class boundary samples; and secondly, performing k-means clustering on the majority class samples and the minority class samples, determining the oversampling weight of each cluster for the minority class samples according to the in-cluster boundary sample density, and selecting a sample point closest to the clustering center for the majority class samples to carry out undersampling, so that the data reaches a balanced state. According to the method, the class imbalance problem is solved to a certain extent, and meanwhile useful information can be prevented from being lost in the sampling process.

Description

technical field [0001] The invention relates to the field of data mining, in particular to a boundary-based mixed sampling method. Background technique [0002] Imbalanced data refers to a data set with an unbalanced distribution of classes. Generally, a class with a small number of samples is called a minority class, and a class with a large number of samples is called a majority class. In many practical application problems, such as network intrusion detection, credit card transaction detection, medical diagnosis and fault detection, etc., the key to solving such problems is to efficiently and accurately identify minority samples. However, the high accuracy rate of many traditional algorithms can be found through analysis only for the recognition of majority class samples, and does not perform well in the recognition rate of minority class samples. [0003] In the solution to unbalanced data at the data level, due to certain defects in both oversampling and undersampling,...

AI Technical Summary

Problems solved by technology

However, in the existing mixed sampling, the role of boundary samples is rarely considered, resulting in poor classification effect

Images