Design And Implementation Of Game Commodity Recommendation System

In recent years,with the rapid development of Internet technology,people are generating data anytime and anywhere.This has led to an explosive growth of data.In this context,recommendation systems have become the main way for users to filter out effective information from massive amounts of data.However,in prccal applications,due to the inconsistency between online service usage characteristics and offline model training usage characteristics,offline characteristics cannot reflect the real state of users,and lack of user characteristics,the model recommendation effect is not good and the user experience is not good.In response to the above problems,this thesis first reviews the development history of domestic and foreign recommendation systems,and summarizes the main problems existing in the current recommendation systems.On this basis,this thesis aims at the inconsistency of online and offline features in the offline framework of the industrial recommendation system,builds a mainstream real-time computing framework Flink recommendation system,and applies it to game product recommendation scenarios.At the same time,in order to test the effect of the algorithm more reasonably,this thesis builds the AB Test online traffic layered experimental framework.Based on this framework,the recommendation system built in this thesis has achieved better recommendation results.The main work of this thesis is summarized as follows:(1)This thesis analyzes the problems of offline sample splicing and uses a real-time computing framework to solve the problem of inconsistencies between the online service usage characteristics and the model training usage characteristics caused by offline sample splicing.This thesis conducts algorithm experiments through online real game product recommendation scenarios and finds that the ARPU(Average Revenue Per User)recommended by the real-time sample stitching model and the model trained with offline sample stitching increase by 10%.In addition,for better business migration,this thesis modularizes the code function,that is,system users only need to write configuration to complete the release of new tasks.(2)In order to standardize the use of features in the entire recommendation system,this thesis designs a feature management module to achieve unified management of features.This module is convenient for system users to add or delete features.At the same time,add default features for non-featured users,and try to solve the problem of user nonfeatures.This default feature module improves the ARPU effect of exposure by 2%.(3)This thesis abstracts system functions to facilitate system users to deploy in different business scenarios.The experiment found that the deployment time of new tasks has been shortened by 50%,while the online deployment process has been standardized to reduce the space for making mistakes.Based on this system design concept,the scalability and maintainability of the system are increased.(4)This thesis builds an online AB Test traffic layering experiment platform and tests the effect of the algorithm as much as possible by means of traffic cutting.In addition,this thesis uses the real-time calculation framework Flink to perform real-time statistics on the effect indicators of the algorithm to make reasonable decisions as soon as possible.In summary,this thesis not only builds a high-availability and high-reliability realtime product recommendation system,but also solves the problem of inconsistent online and offline features through system design,and finally verifies it through the AB Test layered experiment.Experiments verify that real-time sample splicing and realtimeization of relatively fast-changing features can significantly improve the effect of model prediction.