Research On Click-through Rate Prediction Model Of Internet Advertising Based On Deep Learning

Internet advertising is one of the main economic sources of the Internet industry.With the development of machine learning and Internet,the effective use of massive advertising information has greatly increased the commercial value of online advertising.In the data age,as one of the core technologies of advertising delivery system,click-through rate prediction mainly studies the potential relationship between users and advertisements under massive data to achieve accurate advertising delivery.In advertisement click-through rate prediction,data sets are usually sparse.Traditional methods are based on linear models,and features are independent of each other,depending on the artificial feature combination.However,the model of automatic combination feature lacks the deep-seated high-order non-linear feature,and the ordinary deep learning models still has difficulty in training,and does not consider the problems of deep and shallow feature fusion.The goal of this paper is to mine potential user-advertising links through deep learning model for given data,so as to achieve more accurate prediction of advertising click-through rate.The specific research work is as follows:(1)Aiming at the problems of the existing models,this paper proposes an improved Field-aware Factorization Machines Based on Stacked Auto-Encoder model based on Wide & Deep model.Field-aware factorization machines and stacked autoencoder are introduced as wide and deep parts of the model respectively.Compared with the logistic regression of the wide part,the factorization module realizes the automatic combination of shallow features and avoids the problem of a large number of artificial combination features.In the deep part,the unsupervised layer-wise pretraining of stacked auto-encoder is used to extract the high-order non-linear features,and the trained embedding layer parameters are used as the initialization weight of factorization module,which speeds up the convergence speed of the model to a certain extent.Finally,selective fine-tuning is carried out in the overall training of the model,so that the model efficiently integrates the shallow combination features and the deep high-order non-linear features.Experiments show that the model has better prediction effect when the input data are consistent.(2)This paper analyses the characteristics and problems of the experimental data set,and preprocesses the data accordingly.On this basis,starting from the actual business,this paper constructs the historical click-through rate features.Since the user log can not be directly added to the training,this paper constructs the user behavior features from the user behavior log according to the two directions of behavior type and time period.Experiments show that the accuracy of prediction can be improved by adding these two sets of features.(3)This paper designs experiments,analyses and determines the number of nodes and the depth of hidden layer in the greedy layer-wise pre-training process of stacked auto-encoder,and compares the effects of the number of fine-tuning layers of different hidden layers and different hyper parameters on the model.Finally,by comparing the SAEFFM model with the baseline model,it is proved that the convergence speed and prediction effect of SAEFFM model are improved obviously.