Personalized Recommendation Algorithms Considering Rating Behavior Of Customers

With the development of Internet and Communication,information increases sharply.Behind information,huge commercial value is worthy of being explored.However,the burden of users on selecting desired products and services is also increased.This social problem is known as "information overload".Personalized recommender system(RS)has been proved to be an effective measure to alleviate information overload.The basic idea is to use the historical information of users,design recommendation algorithm to extract user preferences,predict and recommend products that users may be interested in,which greatly reduces the user's time consumption in selection.Since rating information is the most direct indicator to reflect user preferences and it is easy to collect.Therefore,most of the recommendation algorithms focus on user rating behavior,but there are still many problems worthy of improvement and in-depth study.It mainly includes the following three aspects:First,research on the natural noise management based on user rating behavior.The rating data provided by users is not completely accurate,and there existed some noisy ratings in RS.The main reason is that users likely are affected by the external environment and the mood in the process of rating,which makes the ratings provided by users have a certain degree of error.Eliminating this kind of noise can improve the performance of the recommendation algorithm to some extent.The existing noise management have some defects,such as inaccurate rating classification,inadequate noise correction and low efficiency.To solve these problems,this paper proposes a new fuzzy approach to manage noise,which can alleviate the defects of the existing methods.Our proposed method introduces the fuzzy set tool into the classification of rating.Since the boundary between the rating categories is not clear,it is more reasonable to use the membership function in fuzzy theory to describe the rating categories.On the basis of classification,a more efficient noise detection and correction method is proposed,and the proposed method is more prominent in the similar noise management.Second,reseach on personalized recommendation algorithm based on rating value.The rating value as the explicit feedback of user preference can reflect the user's preference to some extent.Collaborative filtering as a classic recommendation algorithm,is always limited by the co-rated items when designing similarity measurement.The data utilization rate is extremely low and the recommendation performance is greatly affected.To solve this problem,this paper introduces ?-divergence in signal processing field into the similarity measurement of collaborative filtering,and measures the similarity from the perspective of rating probability distribution,so that the data utilization rate can be increased to 100%.The proposed similarity measurement method based on ?-divergence has greatly improved the prediction error and recommendation performance,and has certain application value.Third,research on personalized recommendation algorithm based on rating preference.Although the rating value as the explicit feedback of user preference,users generally have rating tendency,which indicates that the true preferences represented by the same rating value are still different for different users.Up to now,few literatures focus on the issue.The general idea is to convert the rating value into the preference,and then generate recommendations.In this paper,we notice that the user's preference is always fuzzy,so we use fuzzy tools to transform the rating into interval preference,and design an interval similarity measurement method for collaborative filtering.This method explores the relationship between rating and preference from a new perspective,which has certain enlightening significance.