Research On Optimization Method Of Click Rate Prediction Model In Recommendation System

With the development of the recommendation system,the click rate estimation has been widely applied in the current computing advertising and recommendation systems,and the advertising business is also the core business of most companies.The advertising revenue is the main income of most Internet companies.So major technology companies are looking for ways to improve the accuracy of click-through rate estimates based on actual business needs.At present,the meth od of using the click rate estimation in the recommendation system is more based on the shallow model or the deep learning model,wherein the shallow model directly uses the low-dimensional features or the low-dimensional features constructed by the artificial feature engineering in the feature,and between the feature dimensions isolated from each other,it is impossible to discover the relationship between them.Although the pure deep learning model can achieve nonlinear combination of features at higher latitudes,it also loses the information contained in low-dimensional features.The goal of this paper is to optimize the click rate estimation sorting algorithm by given the recommended system scene data,and to use the shallow model to extract the low-dimensional features and the deep learning model to express the high-dimensional features,and combine the two features.The advantage of mining the relationship between more features can improve the accuracy of the click rate prediction to some extent.Specifically,this paper mainly includes the following three aspects of research.First,this paper starts from the definition of the click rate estimation of the ranking module in the recommendation system,analyzes the distribution and characteristics of the recommended system data set and preprocesses the data set.In addition,this paper based on the understanding of the recommendation system.And its characteristics in practical applications,extracted three different types of features.Secondly,this pap er uses the optimized factorization machine model and deep neural network to predict the click-through rate of the data in the recommended system scenario,and analyzes their shortcomings,and uses the dropout method in the neural network module.To reduce the over-fitting problem that may be caused during the training.Thirdly,this paper adds the user history click skin sequence feature to the model input data to learn the association between the user history click behavior and the item to be predicted,and independently designed the experiment to verify its validity.The optimized click-rate estimation model runs stably on the prediction of the skin recommendation data on the input method homepage,and effectively improves the prediction effect.It can be applied well in the recommendation system click-through rate estimation problem.