Research On CTR Prediction Based On Deep Learning

In recent years,the technology of data collection on the Internet is gradually maturing,the problem of information overload is becoming more and more serious,and recommendation systems are one of the effective tools to solve the information overload.In the industrialized recommendation system,click-through rate prediction becomes an important step in the ranking stage of the recommendation system,while the traditional click-through rate prediction model can no longer meet the current task requirements.With the development of deep learning technology in various fields,it is gradually applied to the click-through rate prediction task of the recommended system.Based on the research and analysis of the existing click-through rate prediction models based on deep learning,this thesis proposes two new click-through rate prediction models.Based on the research and analysis of the existing click-through rate prediction models based on deep learning,two new click-through rate prediction models are proposed.Firstly,click-through rate prediction model based on the xDeepFM.The model mining the combination relation between features from multiple angles makes the expression ability stronger and can realize the full automatic learning.Introducing attention mechanism,the importance weights of features can be dynamically learned according to scene requirements to suppress noise data.The activation function is improved,and the step change point can be flexibly adjusted according to the data distribution,which effectively improves the prediction accuracy of the model.Comparative experiments have shown that the accuracy and bias of the model's predictions are superior to those of deep learning at this stage.Secondly,click-through rate prediction model based on improved xDeepFM and user interests.The model uses the Multi-Head Attention to extract interest factors from the user's behavioural sequence to ensure the evolution of the user's interests.The output of network is a vector of representations containing user interest factors,which can be combined with features combination features to improve the diversity of recommendations in the overall model.Experiments have shown that the model has a higher prediction accuracy than other prediction models.