A New Non-iterative Approach To Kernel Logistic Regression For Imbalanced Data

The classification of imbalanced data widely exists in financial,medical,telecom-munications and other fields,while the traditional classification methods such as deci-sion tree,Support Vector Machines and so on have the problem of inaccurate recogni-tion of minority classes.In this paper,a learning method of non-iterative kernel logistic regression is proposed for severe imbalanced data,which not only reduces the com-putational burden caused by iteration,but also utilizes the basic category proportion information in model training,avoiding the usual way of dealing with imbalanced data such as under-sampling and over-sampling,so that in the case of large-scale imbalanced data,it can be used.It is convenient to model the kernel logistic regression and construct a robust approximate least squares logistic regression classifier.The theoretical study shows that the method proposed in this paper has some good qualities.The simula-tion and empirical analysis show that the proposed method greatly reduces the training time compared with the classical iteration method in the training of kernel logistic re-gression,and performs better in the case of imbalanced data,especially in the case of severely imbalanced data,than the combination of under-sampling and over-sampling classification methods.In addition,the orthogonal table experiment design is used to select parameters in the kernel logistic regression,which greatly reduces the time of parameter adj ustment.