Linear model stable learning method and device

Abstract

The invention provides a linear model stable learning method and device, and the method comprises the steps: carrying out the random resampling of each column of data of an initial matrix of an independent variable, and generating a final matrix according to a sampling result; endowing the first sample label with an initial matrix, endowing the second sample label with a final matrix, and generating a probability binary classifier; and obtaining a new sample weight through a probability binary classifier and applying a probability density ratio estimation algorithm, and adjusting the sample weight according to the new sample weight to eliminate colinearity between independent variables and generate a final linear regression model. According to the method provided by the embodiment of the invention, the purpose of eliminating colinearity between independent variables is achieved through learning of a linear model stabilization method, the fitting precision of the model is improved, andthen stable prediction performance can be kept under variable data distribution.

Description

technical field [0001] The invention relates to the technical field of regression analysis and stable learning, in particular to a linear model stable learning method and device. Background technique [0002] At present, regression analysis is a classic statistical machine learning method, and the linear regression model has been widely used to describe the relationship between dependent variables and independent variables because of its concise mathematical expression and efficient solution process. [0003] However, linear regression will introduce systematic deviations when the model assumptions do not match the real data generation mechanism. This deviation will be amplified indefinitely by the collinearity between independent variables, which greatly affects the fitting accuracy of the model. In addition, traditional machine learning methods are highly dependent on the assumption of independent and identical distribution of data, but in real scenarios due to the unknowa...

AI Technical Summary

Problems solved by technology

[0003] However, linear regression will introduce systematic deviations when the model assumptions do not match the real data generation mechanism. This deviation will be amplified indefinitely by the collinearity between independent variables, which will greatly affect the fitting accuracy of the model.

In addition, traditional machine learning methods are highly dependent on the assumption of independent and identical distribution generated by data. In real scenarios, due to the unknowability of data sources, this assumption is not easy to satisfy, so the prediction performance in real environments is unstable.

Images