Research On Advertising CTR Prediction Model Based On Deep Learning

Click-through rate(CTR)prediction is the process of analyzing log data,extracting valid feature combinations that affect click results of users,and then predicting the click-through rate of advertising products.In view of the shortcomings of using fully-connected neural networks for high-order feature combination,including difficulties in parameter learning,poor learning effect,and lack of interpretability in the interaction process,this paper proposes two networks: Enhanced Factorization-machine supported neural networks(EFNN)and deep self-attention interactive network.The enhanced FNN adds a convolutional feature interaction unit to reduce the difficulty of learning feature combinations,and the deep self-attention interaction network adds a self-attention feature interaction unit to make the model have the ability of learning both explicit and implicit feature interactions.It provides better interpretability on the basis of improving the accuracy of prediction.The main content of the thesis includes the following aspects:First,the domestic and foreign click-through rate prediction models are introduced and summarized according to the structure in this paper.On this basis,the basic knowledge of deep learning and related technologies of CTR prediction are explained in this paper.Second,the structure of several high-performance deep learning models is studied in this paper,and the fully-connected neural network combined with the effective feature interaction units in parallel can improve the prediction accuracy of the model is found.Third,a click-through rate prediction model based on the enhanced FNN which adds a convolutional feature interaction unit to Factorization-machine supported neural networks(FNN)is proposed in this paper.This unit designs a convolution method for CTR data,which can effectively extract feature combinations and reduce the difficulty of learning by only fully-connected neural networks.Experimental results show that the area under curve(AUC)of this model is improved by 0.4243% compared with the original FNN.Compared with the other reference models,this model has also achieved relative improvements of performance.Fourth,a click-through rate prediction model based on the deep self-attention interactive network is proposed in this paper.The self-attention feature interaction unit of the model uses the self-attention mechanism to judge the correlation between features,which can effectively capture the global feature combination,and the interaction process occurs between the embedding vectors of the features,which is an explicit interaction with stronger interpretability.The explicit interaction ability combined with the implicit interaction ability of the fully-connected neural network,which can effectively improve the learning ability of this model.The experimental results show that the AUC of this model reaches 0.7878,which has strong prediction ability.