The Design And Implementation Of Stranger Social Matching Recommendation System

With the rapid development of information technology,we can easily communicate with each other on mobile devices from thousands of miles away,and online stranger social networking has come into being.It has also become a difficult task to find people with similar interests in the vast amount of information on the Internet.As an effective way to filter information,recommendation system has received a lot of attention.The main research of this paper is a recommendation system made for a matching scenario of a stranger social app.The purpose is to make the matching between users more accurate and personalized,and then improve the stickiness of the APP to users.The recommendation system designed in this paper is divided into two phases,"recall" and "sorting".When recommending to users,firstly,the recall stage generates a high quality candidate set,and then the sorting stage finely sorts the candidate set to generate the final recommendation results.Before designing the system,we first analyze the requirements in the actual business scenario,and then design the system architecture and recommendation process accordingly,and introduce the functions,design and implementation methods of each module in detail.The recall strategy was designed with business requirements in mind and Elastic search was selected as the search engine for the recall phase.Source data consists of user information and their historical behavior data.Firstly,in order to standardize the data storage format,the data source types were analyzed and a unified data table structure was designed.The feature list was then determined after analysis of the user profile.For data processing,online data is processed using Flink for real-time statistical calculations;offline data is processed using Hive and its custom UDF functions on a regular daily basis,and the completed offline data is stored on Redis.The recommendation algorithm uses a feature combination hybrid recommendation,which uses a combination of ALS collaborative filtering algorithm and content-based recommendation algorithm.The inner product of the current user vector and the exposed user vector calculated by the ALS collaborative filtering algorithm is used as the rating value of the current user,combined with other user profiling data,the XGBoost model is trained on Spark,and then store the completed training model on an FTP server.For online prediction,the online data processed by Flink and the offline data stored on Redis will be stitched together as feature data,and the model file stored on the FTP server will be used to sort the data by the model for prediction,and the sorting result output by the model will be returned to the user as the final result.Finally,a set of index evaluation system is designed to measure the operation effect of the model online,and the model effect is verified by AB-Test experiment.The results show that the performance of all indicators after accessing the recommendation system is better than the default strategy,and some indicators even exceed the default strategy by more than double.The article concludes with a plan of where the system can be continuously iterated and improved.