Design And Implement Of FreshIndex-based Personalized Recommendation Algorithm

With the rapid development of information technology, various types ofinformation grow explosively, how to effectively extract information which suits theusers’ demand from the huge information ocean has become a problem of goodapplication perspective. By the advancing of studies, the focus of both study andapplication swtches from actively searching to automatically recommandationaccording to the interests of users. Presently, on-line trading is widely accepted,sellers prefer to promote and sell goods through online malls, and personalizedrecommendation technology fits well to e-commerce, which could improve usersatisfaction purchase rate significantly.This project origined from a semantic search and personalized recommendationproject of large electronic business enterprise. The thesis firstly reviewed the currentstatus of information retrievel and recommandation technologies studies, andsummaried few problems of recommandation systems when applies to pracitcalsituations such as cold start and recommendation domain knowledges. Then, throughthe consumption survey of crowds, we found that the crowd is sensitive to theanouncing time of product. Thus based on the correlation relationship, a user freshindex interest model is designed. Fresh index indicates how people feel with theannouncement time of products, for example, it gets the maximum value at the verydate of announcement. In order to make a good match between customer-product foreffectively recommandation, a new concept--consuming index--is defined to describethe average fresh index of products an individual cosumer boughts. Next, based on theuser fresh index interest model, a personalized recommendation algorith is designed,and at the same time, the algorithm is applied to a personalized recommendationsystem of one big electronic business enterprise. Through model analysis andsimulation of fresh index, and the adjustment of algorithm parameters according tothe operation circs, the algorithm is adjusted to achieve a best effect. The system hasbeen under stable operation for six months, with traffic close to one million.Generally speaking, because the universality of fresh index and consumer indexconcepts, the model could help tohandle the cold start and sparse data problems,which would help to reduce the initial data acquisition cost to start therecommendation system, and strengthen the user-item and item-item relationships which whould reduces the diffculty of data retrieveing of cooperative filter fromnearest users, and finaly, improved the purchasing rate of personalizedrecommendation systems. All these characteristics provide a solid support to smalland medium enterprises on the start and operation of recommendations systems.