Research On Advertising CTR Prediction Method Based On Deep Learning

With the development of Internet technology,Internet advertising has emerged as the product of its development.Improving the prediction accuracy of ad CTR is an important task in the field of computing advertising.If the accuracy of CTR prediction is higher,the information provided to users will be more accurate,and the promotion effect of the merchant will be better,which will increase the profit of media platforms,DSP companies and merchants.At present,click-through rate prediction models have attracted the attention of researchers.However,with further research,we can find that there are still problems in the construction of CTR prediction models in the following aspects:Sparse data: Because the number of ads is quite large,only a few ads have rich historical click data.Therefore,the historical click record containing user preferences is rather sparse.The sparse data makes it impossible to accurately calculate the probability that user click on an advertisement,and thus cannot accurately target users based on the content of the advertisement.Scalability: Thousands of ads and users will be involved in the process of advertising click-through rate prediction.The number of ads and users has grown exponentially with the development of the Internet.How to update the click-through rate prediction model and predict the click-through rate of ads accurately has become an important challenge in the forecasting process.Cold start: When a new ad is added,it is difficult for the prediction model to predict the click rate of the ad accurately due to the lack of sufficient historical click data.In addition,when a new user joins,the prediction model lacks relevant information for the user,and it cannot accurately deliver relevant ads to the user.Capturing features: Advertising data features have a highly non-linear relationship.It is important to mine the hidden feature interactions behind the user 's click behavior in theprocess of ad click rate prediction.Therefore,effectively modeling feature interactions becomes a challenge.In response to these problems,researchers have proposed many solutions.In recent years,the rapid development of deep learning technology has achieved breakthrough results in many fields,including speech processing,computer vision,natural language processing,and so on.In essence,the concept of deep learning originates from the research of artificial neural networks.Deep learning forms a more abstract high-level representation by combining the underlying features to obtain an effective representation of data features.Because of overcoming the obstacles of traditional click prediction models and having a more accurate prediction rate,click-rate prediction based on deep learning has gained more attention.Therefore,research on prediction of ad click rate based on deep learning has important theoretical significance and application value.Based on the National Natural Science Foundation of China,this paper conducts an in-depth research on the prediction method of ad click rate based on deep learning for the problems existing in the above existing research work.The main contributions of this paper are as follows:1.Research on CTR Prediction Based on Deep LearningThere are many features that affect CTR prediction,but it is not that the more features considered,the better the prediction effect.To reduce the sparseness of data and to mine the hidden features in advertising data,a method that learns the sparse features is proposed.Our method exploits dimension reduction based on decomposition and combines the power of field-aware factorization machines and deep learning to portray the nonlinear associated relationship of data to solve the sparse feature learning problem.The experiment shows that our method improves the effect of CTR prediction and produces economic benefits in Internet advertising.2.A New Approach for Advertising CTR Prediction Based on Deep Neural Network viaAttention MechanisThe above model can find potential features and fully mine the interaction between features well,but the contributions of different features to the prediction are the same.Therefore,based on the above,this paper uses the attention network to learn the impact of different features on prediction,assigns higher weight to the features that have greater impact on the prediction result,and assigns less impact to the prediction result.At the same time,the data is still subjected to dimensionality reduction processing,and the interaction between features is mined.Experimental results show that using attention mechanism can improve the accuracy of CTR predictions.3.A click-through rate prediction model based on user interestThe previous model only considered some attribute characteristics of the user and the advertisement,and did not take the user interest as an important factor in the click-through rate prediction.Therefore,a click-through rate prediction model based on user interest is proposed.This model can discover the user interest preference from the user historical behavior.Firstly,the high-dimensional sparse feature data is embedded through the embedding layer.Then the user behavior data is processed through the bi-directionally LSTM to discover the user potential interests.At the same time,other feature of the user and the advertisement are captured through the SAE.Finally,user interest features are fully connected with other features to predict the click rate.It can be found through experiments that the proposed model performs better than the traditional click-through rate prediction model that does not consider user interests.4.A Hierarchical Attention Model for CTR Prediction Based on User InterestIn the process of discovering user interest,we finds that user interest has two characteristics,one is that the user interest is diverse,and the other is that the user interest changes dynamically over time.Therefore,this paper proposes a hierarchical attention model based on the evolution of user interests based on the previous model.Specifically,we capture the interest sequence in the interest extractor layer,and the auxiliary losses are employed to produce the interest state with deep supervision.First,we use the bidirectional long short-term memory network to model the dependence between behaviors.Next,an interestevolving layer is proposed to extract the interest evolving process that is related to the target.Then,the model learns highly nonlinear interactions of features based on stack autoencoders.An experiment is conducted using four real-world datasets.Based on the problems in the process of CTR prediction,this paper proposes four methods for predicting the CTR of advertisements.The first method involves the contents of Chapters 2 of this paper.According to the characteristics of the advertising data,the data is reduced in dimension.At the same time,the improved factorization machine model is used for low-order feature interactions.SAE mine higher-order feature interactions to improve the accuracy of click-through rate prediction.The second method involves the contents of Chapters 3 of this paper.It uses the attention network to learn the impact of different features on prediction better.It gives higher weight to features that has greater influence on the prediction result,and it is important to distinguish different features.The third method involves the contents of Chapters 4 of this paper,which discovers the user interest from the user historical behavior,and uses the user interest as an important factor in the click-through rate prediction.The third method involves the contents of Chapters 5 of this paper.It takes into account the user interests while capturing the dynamic changes of the user interests,effectively modeling the evolution process of interest related to the target advertisement,and improving the accuracy of the click-through rate prediction.In summary,this article discusses the problems in the CTR prediction model.In order to improve the accuracy of click-through rate prediction,deep learning technology is used to build models,and the effectiveness of the research work in this paper is proved by theory and experiments.