Research On Key Algorithms Of Ensemble Learning For Click-Through Rate Prediction

With the rapid development and universal of the Internet,the combination of the traditional advertising industry and the Internet makes it possible to advertise online.As an important channel for Internet ads delivering,the real-time bidding system has become highly concerned in academia and industry.In real-time bidding system,research on demand-side platforms is mainly focused on click-through rate prediction,bidding algorithm,and bidding landscape forecasting.Among them,the click-through rate prediction is a primary measure of user feedback and a major basis for bidding algorithms.It has always been a main research direction in real-time bidding system.In recent years,machine learning and deep learning have been used for click-through rate prediction.Researchers have proposed many excellent click-through rate prediction models,summarized which,based on the idea of ensemble learning,and combined with the deep residual network,this thesis proposes a click-through rate prediction fusion model of residual network,and further proposes a click-through rate prediction framework on ensemble learning.The main work and research contents of this thesis are as follows:(1)This thesis summarizes and classifies the CTR prediction models proposed in recent years,which finds that it is a trend of using deep neural networks in CTR prediction models.Residual network is widely applied in the field of image recognition,which effectively solves the problems of vanishing gradient and exploding gradient caused by too many network layers in deep neural networks.Therefore,this thesis combines the structure of residual network and click-through rate prediction model,proposes two single models built upon residual network on the basis of FNN(Factorization Machine supported Neural Network)model,named ResNet1 and ResNet2.(2)Based on the residual network of two single models,this thesis combines the two models through the design ideas of merging models to get ResNet,a fusion model of residual network.The ResNet model contains two different residual network designs,and the Attention mechanism is adopted in the embedding layer to assign importance weights to second-order features,which further optimizes the model's performance.Comparison experiments on two different datasets prove that the performance of ResNet1,ResNet2,and ResNet models surpass all the comparison models.On the Criteo dataset,the ResNet model has improved the AUC index by 0.24% and the LogLoss index by 0.42% compared with the optimal Wide&Deep model in the comparison models.On the Avazu dataset,the ResNet model has improved the AUC index by 0.34% and the LogLoss index by 0.39% compared with the optimal FM model in the comparison models.(3)This thesis also proposes a heuristic ensemble learning framework for integrating click-through rate prediction models.By inputting the prediction results of different click-through rate prediction models for the samples into the framework,the framework automatically calculates the degree of difference between the models,and selects the models with better performance and larger differences for integration,and ultimately obtains the integrated CTR prediction model.The experiments on four datasets of iPinYou prove the effectiveness of the framework.