Classification prediction method and classification prediction device for non-balanced data sets

Abstract

The invention discloses a classification prediction method and a classification prediction device for non-balanced data sets. The classification prediction method comprises the steps of obtaining training sample sets in the non-balanced data sets and an optimal classification result corresponding to the training sample sets; based on a current objective function, classifying the training sample sets to obtain a current classification result; judging whether the current classification result is consistent with the optimal classification result or not; if yes, taking the current objective function as an optimal objective function, or otherwise, based on classifier evaluation indexes, performing performance evaluation on the current classification result to determine a current rewards and punishment function; and according to the current rewards and punishment function, correcting the current objective function to obtain a current corrected objective function, and by taking the current corrected objective function as the current objective function, performing re-classification. The rewards and punishment function is introduced; and according to the current rewards and punishment function, the current objective function is continuously corrected to obtain the optimal objective function, so that accurate classification prediction of the non-balanced data sets is realized.

Description

technical field [0001] The invention relates to the technical field of unbalanced data set classification, in particular to a classification prediction method and a classification predictor for unbalanced data sets. Background technique [0002] Classification predictor is an important information processing technology used in many industries to predict the probability or possibility of the future development of an event. Using a classification predictor to classify the data representing the event can be used in the analysis of many industry data to predict the possibility of the corresponding event. [0003] However, many industry data are typical unbalanced data. Taking the binary classification problem as an example, the characteristic of unbalanced data is that the proportion of one class of data is much higher than that of the other class. Here, the data with a high proportion is called the majority class, and the data with a low proportion is called the majority class...

AI Technical Summary

Problems solved by technology

[0006] However, the classification performance of these two methods for unbalanced data: resampling method and cost-sensitive method has not yet achieved the desired effect

Both of them will increase the classification performance of the minority class to varying degrees, while the classification performance of the majority class will decrease.

The defect of this method is that on the one hand, the proportion of the two types of data in the original data set is changed, so that the resampled data can no longer truly reflect the distribution characteristics of the original data; on the other hand, the classifier is overtrained on the minority data. , so that the training ability of the majority class data is greatly reduced, which leads to the improvement of the minority class prediction performance, while the majority class prediction results are reduced.

However, this method will also reduce the classification performance of the majority class when the cost sensitivity factor of the minority class increases.

Images