Application And Research Of Machine Learning Ranking Model In Personalized Recommendation System

With the development of E-commerce and mobile internet technology,personalized recommendation system has drawn a lot of attention.However,on the base of real-time behavior and history behavior of users,most of recommendation systems represented by E-commerce recommendation system adopt different algorithms to get corresponding result.After filtering and sorting by forecasting score,the recommendation system can obtain the final result.For the result generated by different algorithms,the way that we determine the position of commodities just by forecasting score is so hasty.Also this sorting method can't even reflect the real preference of users and update the result according to users' real-time behavior.Thus,most of the time,the recommendation result is not accurate enough.It is in this context,we propose a sorting recommendation system project.This thesis is based on the mobile platform.According to the user-book scoring information,users' behavior of ordering,reading,and browsing,we extract the features of users,books and user-book,combined with time dimension and book popularity to build multiple machine learning sorting model.And then,we design and implement the sorting personalized recommendation system on the basis of open sourse data framework of Hadoop,Storm and Spark.The system has the following characteristics,first of all,it has the characteristic of real-time,which can provide users with relevant recommendation result fast and efficiently according to the users' real-time behavior,greatly enhance the users' reading experience;secondly,for recommended users,the forecasting score is no longer still,while will be influenced by many factors including freshness,popularity,browsing effect and evaluation of all the users.Undoubtedly,for the recommended sort of the result,these factors can reflect the users' integrated evaluation on the target;last but not least,the system has the characteristic of versatility,which is flexible to a variety of recommendation algorithms and lays the foundation of adding new recommendation algorithms to the system afterwards.After the completion of the design and implementation of the system,we conducted a functional test of the system at first,then we tested the performance of the components in the system,at last,we use the A/B test method to verify the common indicators of recommendation system,like accuracy,AUC and MAP,which is aimed to gurantee the reliability and validity of the recommendation system.At present,the sorting recommendation system described in this thesis has been successfully running in the background of migu reading app,and has brought a good income for the mobile platform.