Research On User Preference Based On Data Mining

With the rapid development of the Internet and mobile Internet,people have become increasingly inseparable from the Internet,which is affecting all aspects of people's lives.However,as the Internet explodes,it becomes a dilemma to find truly useful information for users in the vast amounts of data as people enjoy a wide variety of convenient Internet services,which is known as "information overload".Personalized services,especially recommender systems,quickly locate information that interested to users through technical means such as information filtering,bring a better Internet service experience to users in the ocean of information.User preference extraction technology is the key to personalized service,which determines the personalized service quality of user experience.The traditional user preference extraction technology is too simple either in the data source or in the model,resulting in the low accuracy or lack of robustness of the user preference extraction result,which has some limitations.When evaluating personalized recommendation systems,people tend to focus on accuracy which aims at user long-term preference.There is little research on serendipity which aims at user short term preference.In the background,this thesis digs the issues related to user model and user preference extraction technology.Based on user preference features,this thesis studies the evaluation criteria and strategies of recommender system which emphasizes on serendipity.The main contributions are described as follows.First of all,this thesis has carried on adequate survey on the current research situation of the user preference,summarized the related theory about the existing user preference model and the user preference expression method and compared the advantages and disadvantages of the existing user preference extraction methods.Secondly,based on the existing research on user preference extraction,a multi-model merge user preference extraction method is designed.After tuning the models,which include support vector machine,random forest,gradient boosting decision tree and XGBoost,we use stacking method to merge the four basic models,use Logistic Regression to tune the weight of every single model and combine all the single models by the weights.Compared with single user preference extraction method,the merged method has a percentile improvement effect on the evaluation index ROC-AUC and has a better robustness,which is of great practical significance for future user preference extraction research.Finally,on the basis of digging out the user preference characteristics,focusing on the user's short-term preferences,aiming at the personalized service recommendation system,this thesis proposes a more flexible recommender strategy that takes serendipity and accuracy into consideration based on user preference model in the movie recommendation background,then this thesis proposes a new improved evaluation serendipity of the recommendation system to help the recommender system to mine user short-term preferences,which gives the direction of guidance to enhance personalized services and has a guiding significance for the promotion of personalized services.