Research On Advertising CTR Prediction Model Based On Convolutional Neural Network

With the continuous development and innovation of Internet technology,online advertising formed through the use of Internet platforms has shown great potential and commercial value.The digital smart Internet era brings tremendous opportunities for online advertising.Compared with traditional advertising,online advertising has the advantages of faster spread,wider range of influence,richer and more diverse content,lower advertising costs,and better advertising effectiveness.Therefore,online advertising on the Internet has become an effective way for companies to enhance their popularity and implement marketing strategies to increase their revenue.At present,Internet advertisements mainly include display advertisements,search advertisements,and other types.But no matter what type of advertising,in order to make full use of its convenience,the key lies in the prediction of advertisement click-through rate.Accurately predict the CTR(click through rate)of the advertisement can not only enhance the user experience,but also increase the profit of the advertiser,thereby further affecting the revenue of the Internet advertising platform.Therefore,using historical data to predict the CTR of ads is very meaningful.Although there are a large number of models used to predict the click rate of advertisements,most of them are based on traditional shallow models.For the multi-field category data of advertisements,shallow models often fail to learn the interrelationships between features and cannot learn the underlying patterns of advertising data.At the same time,using the shallow model to predict the CTR of advertisements often requires a large amount of feature engineering to support it,so it will waste a lot of manpower and material resources.In recent years,deep learning models based on deep neural networks,especially convolutional neural network models,have achieved great success in many areas thanks to its powerful automatic learning features.However,the research on advertising CTR prediction based on convolutional neural networks has just started.So,this thesis studies convolutional neural network advertising CTR prediction model.This thesis focuses on the characteristics of various types of advertising data feature categories and the inadequacies of the existing advertising CTR model.Based on the related theories and technologies of advertising CTR prediction model,a convolutional neural network advertising CTR prediction model based on factor decomposition technology is constructed,and the optimization of prediction model is studied.Finally,the effect of the model on predicting the CTR of the advertisement is verified by experiments.The main work of this thesis includes the following aspects:(1)According to the characteristics of high dimensionality and sparseness of advertising data features,the ability of factorization machine to embed features,the advantages of convolutional neural network sharing local connections,and processing high-dimensional data,construct a convolutional neural network ad click rate prediction model that combines factor decomposition mechanisms.The model introduces factorization and embedding at the bottom of the model,embeds the high-dimensional sparse feature data encoded by one-hot into the low-dimensional vector space,and then uses the convolutional pool operation layer of the model to further learn the underlying features of the data.Finally,through a fully connected layer,the deep feature of various aspects is used to output the CTR of the advertisement.(2)Analyzing the characteristics of advertising data and the advertising click rate prediction model based on convolutional neural network constructed in this thesis,introduce the product operation layer to optimize the structure of the click rate prediction model.This thesis solves the problem that the convolutional layer can not model all the cross features between non-adjacent features by introducing a product operation layer on the feature embedding layer.Non-adjacent second-order features are obtained by the inner product operation of the embedding layer feature vectors to facilitate further learning of deep-level feature modes in the convolutional layer.(3)The performance of this model is optimized by using Selu activation function and improved Dropout method.Selu activation function is used instead of Relu activation function to solve the problems of model training instability caused by Relu,gradient disappearance and gradient explosion.Through the improved Dropout method to adapt to the corresponding Selu activation function,the overfitting of the model is effectively avoided.In the experiment process,AUC and Logloss are selected as model evaluation metrics,and using the iPinYou data set to train and test the convolutional neural network advertisement click-through prediction model constructed in this paper.During the training,the back propagation algorithm and stochastic gradient descent algorithm suitable for convolutional neural networks are used to optimize and update the weights.And through comparative experiments for comparative analysis.The experimental results show that the predictive performance of the model based on the network structure is better when the hidden layer number is 2(ie,two convolution and pool layers).The performance of the traditional activation function Relu is better than that of Sigmoid and Tanh.Compared with the other LR,FM,and CCPM three clickthrough prediction models,the combined click-through prediction model of the convolutional neural network ad-computation factor constructed in this thesis has a better effect on the prediction of ad click rate;the optimized model has a certain improvement in predictive performance.In short,the experimental results show that the advertising click-through prediction model constructed in this thesis is effective,and the optimization technology method given is effective.