Improved Ladder Network For Ad Conversion Predictions

With the rapid development of the Internet,searching advertisement has become one of the most important business models in the Internet industry.At the same time,the search ads attracts a large number of experts and scholars for its huge commercial value and research value,and has been widely studied in the academic circle.The conversion rate of search ads,as an indicator to measure the effect of advertising conversion,comprehensively depicts users' purchase intention of advertising products from the perspectives of advertising creativity,product quality and store quality.Accurate prediction of the conversion rate enables advertisers to match the users who are most likely to buy their own products,and enables users to quickly find the products with the strongest purchase intention,thus improving the user experience in the e-commerce platform.The large amount and complex data structure of advertising conversion rate data bring challenges to the task of accurately predicting advertising conversion rate.Logistic regression algorithm has low complexity,and its objective function also fits the task requirements well,so it is widely used in the prediction of advertising conversion rate.However,logistic regression algorithm has high requirements for feature engineering,and needs to spend a lot of time in feature selection and generation stage.Semi-supervised learning algorithm can effectively extract features and reduce the time of feature engineering,and can be directly applied to the task of advertising conversion rate.Based on the above background,this paper takes semi-supervised learning algorithm ladder network as the theoretical basis,and focuses on the task of advertising conversion rate to conduct the following research:1)analyze the ladder network of semi-supervised learning algorithm and improve its network topology,so as to enhance the expression ability of the model.In this paper,the shallow network is introduced in the encoder part of the ladder network.The number of neurons and the number of layers in the shallow network are small,and the expression ability of the model is improved while the algorithm complexity is appropriately increased.2)regular terms are added to improve the optimization function of the ladder network,so as to improve the generalization ability of the model.Then,according to the characteristics of advertising conversion rate data,the dimensionality of high-dimensional features is reduced through the embedded layer,and features are extracted through different networks according to the characteristics of different types of features,so that the model can better meet the task requirements.3)Finally,through the experiment on ali mama advertising related data,the model has been proved to have a good improvement in accuracy.Then,experiments are carried out on sample data from large-scale data sets with different sampling proportions.The results show that selecting the appropriate sampling proportion can accelerate the training speed of largescale data sets without reducing the accuracy of the model.