A Method for Click Conversion Prediction in Sparse Feature Scenarios

Abstract

The invention discloses a method for predicting click conversion in a sparse feature scene, and relates to the field of click conversion rate prediction. A neural network-based CTR model is established. This CTR model adopts a new neural network structure, which can not only learn the interaction information of low-order features, but also obtain the interaction information of high-order features. In addition, the mutual information of low-order features is not limited to low-order features of linear relationships, but also includes low-order features of nonlinear relationships. In order to prevent the problem of gradient update caused by too deep network layers, we choose to increase the residual network structure to optimize our algorithm model.

Description

technical field [0001] The invention relates to the field of click conversion rate prediction, in particular to a method for click conversion prediction in a sparse feature scenario. Background technique [0002] Click-to-conversion rate, or CTR for short, usually refers to the ratio of users who click on a particular link to the total number of users who view a page, email, or ad. It is usually used to measure the success of a website's online advertising campaign, as well as the effectiveness of email campaigns, and is one of the core criteria for Internet companies to allocate traffic. It can be seen from this that the higher the click conversion rate, the more successful the advertising campaign, the more popular the product, or the better the product sales. [0003] Based on this background, CTR prediction has always been a hot field for machine learning applications, but it is also a hot problem. Because, in the field of e-commerce, the behavioral features of users a...

AI Technical Summary

Problems solved by technology

Because, in the field of e-commerce, user behavior characteristics are usually sparse and huge, which brings difficulties and challenges to modeling

The early modeling method combining artificial feature engineering and logistic regression (LR for short) required a lot of manual processing. The required talents not only need to have an understanding of business and industry, but also have high experience requirements for algorithm processing. Good or bad often depends on the effect of manual processing of special diagnosis; based on this further, the modeling method of combining gradient boosting tree (GBDT for short) and LR requires a lot of manual processing, but due to its interpretability and GBDT for false The improvement of the weight of the example has greatly improved the accuracy of the prediction calculation; then, with the popularity of the neural network, the method of modeling CTR through the neural network has also gradually emerged, and the method of learning user behavior characteristics through the neural network The performance of the model has been greatly improved. For an Internet company with a large user base, a 1% increase in CTR will bring incomparable benefits, not to mention the neural network-based CTR model. The improved CTR is far from More than 1%

Images