Research On Forecasting Algorithm Of Advertising Conversion Rate Based On Multi-model

The development of online advertising has become the main business model and core profit method of major Internet companies.It migrates traditional marketing methods to online.In a measurable,attributable,and directional manner,it creates a tripartite win-win situation that maximizes advertisers' marketing effects,maximizes platform revenue,and maximizes user experience,and fundamentally improves marketing efficiency.With the development of performance advertising,conversion rate estimation determines whether the platform can recommend advertising to the most interested users in the most suitable position,so it has become a very critical part of the advertising system.The accuracy of advertising conversion rate estimation is affected by many objective circumstances such as users,advertisers,advertising spaces,apps,etc.At the same time,the imbalance of positive and negative samples caused by the sparseness of conversion events,and the uncertain error flags caused by third-party return data have brought greater challenges to the accuracy of the estimation.This paper takes the traditional ctr estimation method as an inspiration,focuses on the feature processing method of massive data mining information,and conducts indepth exploration around the GBDT algorithm.On the basis of optimizing a single model,based on the cognition of integrated learning methods,a new idea of advertising conversion rate estimation is proposed in the way of model stacking and integration.This subject is researched and practiced under the guidance of the above ideas,and the final work content and innovative results are obtained,which are mainly reflected in the following three aspects:(1)Starting from the method of acquiring online advertising data,and aiming at the characteristics of large-scale and high-dimensional advertising data,a feature mining method is proposed around user clicks and installation behaviors,and the word2 sec method is proposed based on the similarity of users installing apps.The positive effect can be obtained after verification.(2)Based on online advertising cold start and ratio comparison problems,an innovative feature correction method,namely Bayesian smoothing,is proposed.It has been verified that feature correction can significantly improve the training accuracy of the model.(3)In-depth study of traditional GBDT algorithm and new XGBoost,Light GBM algorithm,and analysis of the effect in practice.Based on the cognition of integrated learning methods,XGBoost and Light GBM are used as the basic model for stacking and fusion,5tand designed two more complex integrated models: XGB?S?LGBM model and LGBM?S?XGB model.Through the performance comparison with the traditional single model,the practical results are demonstrated.In this thesis,the loss function logloss is used as the model's effect evaluation index.After verification of the real advertising data of an Internet company,the XGBoost and Light GBM models solve the problems of high time cost and easy overfitting in the traditional GBDT model.On the basis of a single model,the XGB?S?LGBM and LGBM?S?XGB models proposed in this article further improve the accuracy of model prediction.Therefore,it has a significant advantage on the subject of advertising conversion rate estimation.