Improvement And Application Of Collaborative Filtering Algorithm In Recommender System

With the in-depth development and application of Internet technology,a vast amounts of information on the Internet is generated and distributed everyday,and their speed is faster than the speed of user's responses.We lack a kind of efficient and automatic solution to help users to pick out the items they might be interested in from the massive information under the condition of limited field of user knowledge and unclear demand.In this situation,the Recommender System arised at the historic moment.Recommender System using the algorithm to recommend interesting items to the users.In the common recommendation algorithm,the collaborative filtering(CF)algorithms was studied earliest and used widely.CF is a high-precision algorithm because it's based on “Collective Intelligence”.But the traditional CF algorithm still has some problems,such as inconsistent rating standard?migration of the user interest and low data utilization.In this thesis,we will introduce some related concept to solve these problems,such as time dimension information and dynamic mixing factors.In this thesis,the main work has the following four aspects:(1)This thesis studies the theory of Recommender System.First of all,this thesis illustrates the valuation and composition of Recommender System,and points out that the recommendation algorithm is the core of the Recommender System;Secondly,this thesis expounds the common recommendation algorithm in the Recommender System,and then emphasises on the theory of collaborative filtering algorithm.Finally,this thesis analyze the basic principles of the traditional collaborative filtering algorithm based on user(UBCF)and collaborative filtering algorithem based on item(IBCF),and then compares the two kinds of tranditional collaborative filtering algorithm's advantages and disadvantages.(2)This thesis introduces time dimension concept to improve the accuracy of collaborative filtering algorithm similarity calculation.Firstly,this thesis analyzes the objective phenomenon which is existing in the application of recommender system that is user's rating standard?item's rating standard and user's interst are changing with time every monment.Secondly,this thesis puts forward user's recent active time,items' s recent popular time,user's rating standard and their mathematical modeling;Finally,this thesis improves the similarity weigth calculation formula of UBCF and IBCF algorithm,and then we get the improved similarity weight calculation formula.(3)This thesis proposes TUBCF algorithm,TIBCF algorithm and hybrid TIUBCF algorithm which are based on time information.First of all,this thesis uses improved similarity weights formula to proposes TUBCF algorithm and TIBCF algoritm which are based on time information.Secondly,this thesis points out that single collaborative filtering algorithm still has the problem of low data utilization.So this thesis from the perspective of recommendation results deversity,I combined with weighted hybrid methods to design dynamic mixing factor which is used to mix the TUBCF algorithm and TIBCF algorithm dynamically,and then we get hybrid algorithm——TIUBCF.At last,this thesis used MoiveLens dataset to do the off-line experiment.In this experiment,I compared the prediction accuracy(MAE)of UBCF?IBCF,TUBCF,TIBCF and TIUBCF algorithm to prove that TUBCF,TIBCF and TIUBCF algorithm has certain promotion in the accuracy of recommendation.(4)This thesis is an attempt to apply the recommendation algorithm,I implement a Recommender System of medical electrical suppliers.First of all,this thesis analyzes the demands of medical electrical suppliers Recommender System.Secondly,this thesis illustrates the explicit and implicit user's behaviour data's collection scheme,and then introduces the relevant recommendation module.At last,Through the online small flow experiment of home page recommendation module's click-through rate and conversion rate to prove that the profit that hybrid recommendation algorithm brings.