Online adjustment method of multi-classifier, online adjustment device, storage medium and electronic device

Abstract

The invention relates to an online adjustment method and device of a multi-classifier. According to the online adjustment method and device of the multi-classifier of the invention, after the multi-classifier is adopted to classify online data currently to be processed, the classified data are not directly used to adjust the multi-classifier, instead, the data are cached until various categories of cached data meet a preset quantity condition (the quantity difference value of any two categories of data does not exceed a first predetermined threshold value), and then various categories of datathat meet the preset quantity condition are adopted to adjust the multi-classifier. According to the method and device of the present invention, the various categories of data which have been classified are cached, and the various categories of cached data that meet the preset quantity condition are adopted to adjust the multi-classifier; and therefore, the distribution balance of the quantity ofthe various categories of data which are adopted to adjust the classifier model can be ensured to a certain extent, and therefore, the performance fluctuation of the classification model in a multi-classification problem can be effectively decreased, and the problem of possible inaccurate classification of the classification model can be solved.

Description

technical field [0001] The invention relates to the technical field of data processing, in particular to an online adjustment method, device, storage medium and electronic equipment of a multi-classifier. Background technique [0002] In the field of machine learning, the classification method is a method of inputting the classification model of the data training data that has been marked into the corresponding classification algorithm, and then using the trained classification model to predict the new data category. Among them, the training methods of the classification model include offline training and online training. [0003] At present, the online classification methods corresponding to online training include online gradient descent (OGD, Online Gradient Descent) and stochastic gradient descent (SGD, Stochastic Gradient Descent), etc. However, these online methods have the disadvantage of being difficult to generate sparse solutions, and then predict In the process o...

AI Technical Summary

Problems solved by technology

[0003] At present, the online classification methods corresponding to online training include online gradient descent (OGD, Online Gradient Descent) and stochastic gradient descent (SGD, Stochastic Gradient Descent), etc. However, these online methods have the disadvantage of being difficult to generate sparse solutions, and then predict In the process of classifying new data, there will be problems of high memory usage and time complexity. In view of this, FTRL (Follow The Regularized Leader Proximal) for stochastic gradient descent is proposed in this field to solve the problems existing in the stochastic gradient descent method.

However, in practical applications, the multi-classification model based on this method (ie FTRL) tends to have large performance fluctuations, which may lead to inaccurate classification of the multi-classification model.

Images