| Click-through rate(CTR)of advertising is a key link in Internet computing advertising.The accuracy of CTR estimation directly affects the advertising revenue of Internet companies.The use of machine learning algorithms in the advertisement recommendation system can effectively increase the click conversion rate of advertisements,which can effectively increase the profits of Internet companies.However,some Internet companies still use logistic regression as the main tool for their CTR prediction.In order to capture the nonlinear relationship between data,we need a lot of artificial experience features.This process takes a lot of time,but with limited artificial capabilities,it is difficult to fully excavate the nonlinear relationship.With the same click-through rate estimation problem,it is difficult to popularize a scenario,and relying on artificial features is not smart enough.Therefore,there have been attempts to improve the model in the industry.This article begins with the logistic regression and introduces XGBoost,factorization machine and field-aware factorization machine in turn.We referred to the ideas of the blending model of GBDT and logistic regression put forward in a Facebook's paper.Using GBDT's non-linear fitting ability,the leaf nodes are used as the input of logistic regression to enhance the nonlinear ability.The innovation of this article lies in the further improvement of this method,using an improved version of the GBDT model XGBoost to increase the accuracy of the model.At the same time,we use the improved model factorization machine of logistic regression to further enhance the nonlinear fitting ability of the model.At the same time,we use the word vector to represent the ID feature,which greatly improves the accuracy of the model.The blending model of XGBoost and factorization machine used on the click dataset exceeds any single model on the AUC index and achieves better results,which verifies the effectiveness and feasibility of the method. |
PDF Full Text Request