Research On Advertising Click Prediction Based On Deep Learning

The accurate prediction of the click event of the advertisem,ent is one of the important factors of the real-time bidding algorithm,which is related to the cost and benefit of the advertiser.In the current online advertising business scenario,the data scale is huge and has characteristics of high dimension,nonlinearity and sparsity.How to design an algorithm to reduce iteration is a key factor to consider in the optimization algorithm.At the same time,deep learning also provides an effective means for learning and predicting large-scale nonlinear features.In view of the above problems,this paper will be divided into two parts to carry out research,work on the ad click prediction algorithm.(1)Based on the research of FTRL(Follow-the-regularized-Leader)optimization algorithm,,the paper proposes a factorization machine based on FTRL optimization algorithm.In the stage of feature extraction,the paper replaces the artificial feature extraction with the feature extraction method based on GBDT(Gradient Boosting Decision Tree)algorithm.The FTRL algorithm is improved on the b,asis of the gradient descent algorithm,and a mixed regular term is added to prevent over-fitting.The FTRL algorithm improves the sparsity while maintaining the high accuracy of the gradient descent algorithm.The factorization machine can better learn the relationship between the different feature components,which is more effective than the traditional logistic regression.The comparison experiment results show that the improved algorithm effectively improves the prediction accuracy of the advertising click event.(2)Aiming at the time-consuming and laborious extraction of artificial features and the shallow model can not fully consitder the nonlinear relationship between data,the paper proposes a hybrid neural network model based on CNN(Convolutional Neural Networks)-LSTM(Long Short Term Memory).The sparse connection and weight sharing characteristics of convolutional neural networks can automatically extract high-impact features and avoid complex artificial feature extraction.Compared with the recurrent neural network,the LSTM neural network has four characteristics:input gate,forget gate,output gate,and cell state.The LSTM neural network overcomes the problem of gradient explosion or gradient disappearance during network training,and classifies it by time sequence.Experimental results show that the improved hybrid neural network is more efficient than the single-structure neural network.