Method and system for training distributed CTR (Click To Rate) prediction model

Abstract

The invention relates to the field of big data machine learning, and discloses a method and a system for training a distributed CTR (Click To Rate) prediction model. Logistic regression calculation can be carried out under a distributed parallel framework, parameters and data can be distributed to multiple primary clusters for being processed; the primary clusters can be further subdivided into multiple secondary clusters, a MapReduce framework is adopted by the secondary clusters in each primary cluster, the gradient can be calculated by Map, and the iteration direction and the step length can be calculated by Reduce; and a task backup mechanism can be adopted in the secondary clusters. According to the method and the system, disclosed by the invention, the sample number and the characteristic parameters, which can be processed, in the logistic regression calculation can be increased, the integrated running efficiency of the system can be simultaneously increased, the calculating and processing time can be shortened, and the phenomenon that the whole calculation is interrupted as a node is in massive failure can be prevented from happening.

Description

technical field [0001] The invention relates to the field of big data machine learning, in particular to a method and system for training a distributed CTR prediction model. Background technique [0002] Logistic regression is generally considered to be a good machine learning method for classification, especially in the aspect of object click-through rate prediction, such as the click-through rate (CTR) prediction of accurate content (such as advertising placement / media recommendation, etc.) Effect. Studies have shown that increasing the number of training samples, increasing the number of features or model parameters will improve the accuracy of classification / click rate prediction, but at the same time, there are stricter requirements on the speed of processing. [0003] In the prior art, in order to achieve the goal of accuracy and speed, there are currently two main methods, one is to use GPU, and the other is to use parallel and distributed processing. [0004] The u...

AI Technical Summary

Problems solved by technology

[0004] The use of GPU can improve the processing speed, but its disadvantage is that the processing speed will not be significantly improved when the training requires a large amount of memory; and when the number of training samples or parameters is reduced in order to improve the speed, it will be leading to a decrease in forecasting accuracy

[0005] Parallel and distributed processing can process a large number of parameters and training samples, but its disadvantages are: the current processing method is mainly concentrated on the linear convex model, first of all, distributed gradient calculation is required, but due to the inconsistent processing speed of each node , so the synchronization requirements have to be relaxed, and the gradient update will not be timely, which seriously affects the speed of processing completion

Images