Research Based On Wide And Deep Model In CTR Prediction

Online advertising revenue accounts for much of the Internet revenue.In online advertising,click-through rate prediction is an important research content.From traditional machine learning algorithms to deep learning technologies,researchers are also constantly trying to apply some new frameworks to click-through rate prediction,and exploring technologies in the fields of computer vision and natural language processing to make them work in click-through rate prediction.This dissertation summarizes the research results in the field of click-through rate prediction in recent years,and finds that some wide and deep models based on feature combination have achieved good results in feature interactive,but they lack the use of the user's historical behavior information.The historical behavior sequence usually implies some important information,such as the evolution of user's interest.Under this background,this dissertation improves the DeepFM model by studying the user's historical behavior modeling in the field of click-through rate prediction.The content of the dissertation includes:First,to summarize and classify the click rate prediction models in recent years,mainly divided into click rate prediction models based on feature combinations and user behavior sequences.Through investigation and research,it is found that there are usually two ways to process the user's historical behavior sequence,one is based on the recurrent neural network,and the other is the pooling operation,which usually uses the attention mechanism.In addition,related technologies are introduced,and the application of deep learning technology and attention mechanism to the click-through rate prediction problem is elaborated.Second,a wide and deep model based on recurrent neural network is proposed.This dissertation takes advantage of the recurrent neural network in processing sequence information,and integrates it into the wide and deep model to make up for the problem that the original wide and deep model lacks the ability to learn user sequence features.A wide and deep model based on the attention mechanism is proposed,which uses the attention mechanism to model the user's historical behavior and explore the influence of different historical behaviors of the user on the current behavior.Data augmentation technology is used to deal with the case where the length of user behavior sequence is too short.In the way of introducing the user's historical behavior sequence,we tried two ways of using input layer introduction and top layer introduction.Finally,design experiments to verify the effectiveness of the models proposed in this dissertation.First of all,this dissertation selects the commonly used models in recent years as a comparison.The experimental results show that the models proposed in this dissertation have improved in AUC,accuracy and logloss indicators.Secondly,the article also compares and analyzes the experimental results of two different methods of user behavior sequence processing,and draws the conclusion that the improved method based on the attention mechanism is superior to the recurrent neural network in the ability to process noise information.The history information connected in top layer outperforms connected in input layer.Finally,this dissertation also explores the effect of different user behavior sequence lengths on the experimental results.It is found that the attention-based approach when processing user behavior sequence information increases the effect as the length increases,RNN increases first and then tends to be smooth.