Research On Recommendation System Algorithm Based On User,product And Context Features

Today,with the booming development of the Internet,online advertising,news,shopping,entertainment and so on are closely related to People's Daily life.However,the mobile Internet produces a large amount of information,and the amount of information people browse is limited.How to select the target data among the complex and diverse information becomes a challenge.Recommendation system has become the mainstream technology of mobile Internet by guiding people to shopping,news browsing and entertainment.Improving the accuracy of recommendation system has become a key topic to solve people's selection difficulties.In order to improve the accuracy of CTR prediction,it is necessary to mine effective information which can guide the model to predict CTR from the huge historical behavior records of users.Therefore,this thesis aims to build a new model to improve the estimation accuracy of CTR.To solve the feature combination problem of recommendation system,an improved CTR prediction model is proposed based on Deep FM model of recommendation system by making full use of user,product and context information.The specific research work is as follows:First,in order to adapt the data to the model and correct or delete records that do not apply to the model,the data is preprocessed in this thesis.The methods of data preprocessing include deleting unique attributes,solving missing values,outlier processing,feature coding,data regularization,selecting main components of feature analysis,etc.In order to reduce the computational cost and transform the data into features that can better represent the potential problems of the prediction model,the thesis conducts feature engineering on the data of the data set.Feature engineering processing is divided into two steps,one is feature selection,the other is feature digitization.The category variables are digitized using unique thermal coding to obtain a form that is easy to be used by machine learning.Secondly,this thesis studies CTR predictive neural network(Deep FM)and adaptive factor decomposition network(AFN)based on factor decomposition machine,and innovatively proposes a factor decomposition machine neural network integrating adaptive factor decomposition network,called AFN+Deep FM model.Recommendation system data are often high-dimensional,multi-domain,sparse,multitype,and rarely associated,while matrix decomposition method in AFN+Deep FM model can fully mine context information for learning.On the one hand,the Deep learning module of the Deep FM model can find the features with hidden relations from the known features and carry out high-order feature interaction with them to improve the model prediction effect.AFN model in front of the feedforward neural network,on the other hand,add a logarithmic transformation layer,adaptive to the characteristics of the input interacts the output characteristics of different order combinations,feedforward neural networks for different order features combination for higher order interactions,AFN model in a different order of eigenvector combination mining for more information,to hit the guidance of the forecast ability stronger.Therefore,the AFN+Deep FM model makes up for the defect that the factorization machine(FM)can only realize the first-order and second-order low-order feature interaction,and mines more information to improve the performance of the fusion model.Finally,the proposed AFN+Deep FM model is verified by experiments.The model was trained and learned on Criteo data set and Avazu data set.Finally,AUC value was improved compared with other mainstream models,while Log Loss value was decreased compared with other models,which proved the effectiveness of AFN+Deep FM model.In summary,the AFN+Deep FM model proposed in this thesis not only performs loworder feature combination,high-order feature combination for input features,but also performs high-order interaction for feature combination of different orders,fully mining information and improving the accuracy of CTR prediction of the model.