The Application And Research On Theory Of Boosting In Recommendation Algorithm

In recent decade, because of its cheapness and convenience, electronic commerce hasbecome a popular way of shopping increasingly. It includes substance commodity、virtual commodity and service，such as subscribing of bell ring, online cinema.Although the quick development of electronic commerce has brought great conveniencefor ordinary people, in the meantime facing so much commodity, customers are trappedin a sea of commodity information，specially some customer without clear target oftensurf to scan the commodity information, but too much information will make customertired when selecting, ultimately transaction will mot be concluded. So how to providethe customer personal commodity information is a way to boost the transactions andsolve the present electronic commerce market bottleneck.In this thesis, first we introduced the purpose and basic principles of therecommending，such as some traditional recommending methods such as collaborativefiltering and matrix factorization, and their advantages and disadvantages. The sellingrecords of Netflix are used as experiment data, then compare the performance of somealgorithms. Then the ensemble learning algorithm is introduced，particularly boosting isselected as main frame of the algorithm in this thesis. According to the comparing ofexperiment result，two recommending methods are considered as the weak methods，finally a efficient recommending algorithm is obtained, the advantage of the algorithmin this thesis is proved by experiment.The main work is reflected in the following aspects:(1) Give a summary of some current popular recommending algorithms and anintroduction of their basic principles. This thesis compared the experimental results ofvarious recommending algorithms on the Netflix.(2) The ensemble learning is applied in several fields，performance can be better. So it isregarded as the frame of algorithm this thesis. Finally two recommending methods are selected as the weak methods in boosting.(3) According to experiment，the main parameter values are determined, finish theensemble of two recommender algorithms.(4) We have proposed several future possible improvements on this subject.