Algorithm Research On Accuracy And Real-time Improvement In Real-time Recommendation

Thanks to the rapid development of Internet technology,the data on the network exponentially increased.People are surrounded by massive amounts of data and find it increasingly difficult to find information of interest.In the situation of "information overload" recommend system came into being.A single recommendation model has various limitations,and researchers often use some combinatorial methods to blend multiple recommendation models,using the advantages of multiple models to compensate for the shortcomings of single model.In addition,the traditional recommendation model needs to update the model periodically.Every time when the model computing recommendations relys on the historical data before last update,so it is difficult for real-time recommending.However,recommended system problems in real-life is more likely based on short-term data,items in such application scenario usually has short timeliness,so real-time recommendation is becoming more and more important.The main research works are as follows:(1)This thesis studied the commonly used model combination methods in recommendation system,and propose a hybrid multi-model collaborative algorithm which applies the concept of "virtual neighbor item".The algorithm uses the idea of collaborative filtering to combine multiple models to overcome the limitations of the traditional item-based collaborative filtering,which can effectively improve the recommendation accuracy.Matrix factorization and Ristricted Boltzmann Machine are two relatively good single models,so this thesis mainly chooses these two models and some of their extended versions to blend.In the experiment,four single modles are optimized through parameter tuning,then we compared their results and choose two better models to blend by multi-model colleborative filtering algorithm.Finally,this thesis compared the results with other model combination methods.(2)This thesis proposed a real-time recommend algorithm based on user's action weight.The users' actions are modeled through the time window.Take advantage of the time dimension information of the users' action data,different weights can be assigned to different actions.The greater weight of action near the current moment,the greater impact on the recommendation results,so that the recommendation results can reflect more relevance with user's current actions.Finally this thesis verified the real-time improvement of the recommendation results through comparative experiment.(3)This thesis investigates three distributed frameworks including Flume,Kafka and Storm.A real-time recommend model was designed based on these three frameworks.The algorithms proposed last two chpaters was also applied in the model.This thesis implemented the real-time recommend model and studied the effect of the recommendations in practice and verified the model can provide real-time recommendations for users.