Research On Advertisement Click Through Rate Prediction Based On The Integration Of ADIN And FM

With the vigorous development of the Internet industry,online advertising has become the main channel for Internet companies to make profits.It has become a crucial measurement task to accurately predict whether the advertising content of a given group is likely to be directly clicked by users of a specific target group based on related information such as direct clicks,browsing,and purchases of group users.Advertisers usually rely on click-through rate(CTR)to distribute traffic revenue.This thesis focuses on the specific characteristics of the various types of prediction of advertising click data characteristics,and the current shortcomings of the existing Internet advertising data click rate characteristics prediction data model,combined with the existing Internet advertising data click rate characteristics prediction data model and related foundations,the study of theory and prediction technology has built a new type of Internet advertising click feature prediction data model that uses Factorization Machine(FM)as a component of the neural network,and has carried out a systematic analysis of the structure of this prediction model.This dissertation optimized the design research,and finally conducted experiments and analysis by comparing the prediction effects of several models on the advertising data set.The main research contents of this thesis are as follows:1)Combining the existing popular types of click-through rate prediction models derived from neural networks,a novel model is proposed:Attention-based Deep Interest Network(ADIN).This model designs a local activation unit,according to the user's historical behavior and a given advertisement,the local activation of the user's historical behavior characteristics,in order to adaptively learn user interests.In addition,an attention mechanism is introduced in the part of user historical behavior analysis,and then the contribution of different user behavior characteristics to the prediction results is distinguished before the characteristics are input to the hidden layer.2)According to the high-dimensional sparse structure characteristics of the original Internet advertising click data characteristics and the processing ability of FM to process the low-dimensional advertising click data characteristics,ADIN uses neural network weights to share local data connection,and the technical advantages of processing and collecting high-dimensional advertising data.A predictive analysis model of Internet advertising click-through rate data that integrates ADIN and FM is constructed.The prediction model first introduces an embedding layer at the bottom of the network,maps the original high-dimensional sparse vector to a low-dimensional dense vector of the same length through calculation or table lookup,and then inputs the sparse features after the embedding layer to FM and ADIN respectively.Finally,the stacking network integrates deep and shallow features to output the prediction results of online advertisement clicks.3)In the activation function part of ADIN,the PReLU activation function is innovatively used,and the adaptive activation function Dice is introduced.The PReLU activation function solves the problem of the conventional activation function in the optimization of the ADIN and uses the gradient descent method to optimize the weights.When encountering problems such as gradient disappearance or gradient explosion,the Dice activation function increases the time complexity of the model while effectively preventing the ADIN from overfitting.The experiment is based on Amazon,MovieLens,and Alibaba advertising click datasets to conduct model training and click results testing on the ADIN and fusion structure model constructed in this thesis.The general AUC index and logarithmic loss index of deep learning are used as the evaluation of model performance.In the training of the ADIN,the fusion model and the four major comparison models,the popular mini-batch stochastic gradient descent algorithm in deep learning is used to update the weights in batches,and the model performance is analyzed by comparing the four popular models.The experimental results show that the ADIN has the best prediction effect in the three datasets when the number of hidden layers after splicing and smoothing is 2 to 3.The training effect of the PReLU activation function in the ADIN model is significantly higher than that of Sigmoid and Tanh.Compared with the other four comparative models of LR,FNN,Wide&Deep and DeepFM,the prediction performance of the ADIN constructed in this thesis has been significantly improved in the three datasets,and the prediction effect of the model combining ADIN and FM is better than that of the ADIN.