Research On Influence Factor Of E-Commerce Collaborative Filtering Recommendation Quality And It’s Improvement Mechanism

Collaborative filtering recommendation techniques is the most mature andsuccessful one in E-commerce recommendation system currently, but it has sparsity,cold start, expansibility problem, which hindered the recommendation quality andefficiency of recommendation for further development. This paper start from thecurrently experiencing problems of E-commerce collaborative filteringrecommendation quality and efficiency, and analyse E-commerce collaborativefiltering recommendation system from three angle, which is ensuring the integrity ofbasis evaluation data, the accuracy of recommendation method, the suitability ofcomputational complexity, then points out the reasons of bottleneck in collaborativefiltering recommendation system, and puts forward the improvement mechanism.The popular current recommendation system found users shopping preferencesbased on user’s project direct score or evaluation data, then applying arecommendation algorithm to excute recommendation to the user. Because some usersdon’t want to evaluate project which result in missing data, leading to the sparsityproblem of collaborative filtering recommendation. Aiming at this problem, this paperputs forward using SOM neural network to cluser users with similar shoppingpreference based on user evaluation data, based on the similarity in the same usercluster, RBFN (radial basis function neural network) is further applied to smoothingprocess to obtain the user’s missing evaluation value. Neural network clustering andprediction effectively eliminates the data sparsity problem of basis evaluation data,reduce the impact of integrity of basis evaluation data to recommendation system.Because of cold start problem, new registered users and products cann’t berecommended in collaborative filtering Recommendation System. To new userproblem, this paper will mining association rules in "Product Taxonomy" which isgenerated based on Vague set theory, and excute recommendation based onassociation rules. Association rules mining and recommendation solved the cold startproblem effectively, as a necessary supplement of collaborative filteringRecommendation, it ensured the accuracy of recommendation. At the same time,based on the traditional user-based collaborative filtering recommendation, this paperproposed the most similar users and the most unsimilar users were clusteredrespectively, and Difference sets of preference product between two clusters wererecommended as a final recommendation to improve the accuracy ofrecommendation.Scalability problem which is caused by system computational complexity, thispaper proposed computational complexity should be considered from datadimensionality reduction and calculation method optimization. This paper proposed akind of "product classification method" based on the vague sets theory, productfeature is extracted and represented with Vague value, and items similarity iscalculated through the Vague value similarity formula. Thus, items were classifiedaccording to the similarity between them, and "Product Taxonomy" is generated. Thismethod can more accurately represent similarty between items, so that similar item clustering will be more accurately. Baded on "Product Taxonomy","seeds" will bepreset based on user’s interest, and calculation and recommendation is implementedin "seeds". Project preclassification effectively reduces the data dimension which needto be calculated, thus reduces the computational complexity effectively. At the sametime calculation method optimization is implemented through combination of"computation offline" and "recommendation online ", which means data pretreatment,similarity calculation, clustering calculation is excuted in offline stage and onlinephase only implement results recommendation, thus give full play to the performanceof the server,so as to ensure the expansibility of system.