| With the rapid growth of the Internet,information channels are intercommunicated,and the volume of data has increased exponentially.Taking the e-commerce field as an example,the number of merchandise entries is in the billions,so user's time is spent heavily on searching and filtering.Excessive information is far beyond the scope of user processing,resulting in user information burden.Therefore,how to alleviate information overload is a hot topic of current research.Collaborative filtering(CF)is one of the most in-depth and mature recommendation technologies,and it is an efficient means to deal with information overload.Similarity measure is the core of recommendation algorithm including collaborative filtering,which greatly affects the accuracy and performance of the recommendation.Many of the previous similarity measures are generally subject to the following issues.(1)Typical similarity measures only use co-rate items,which result in low utilization of data,and have poor performance in high data sparsity environment.(2)It is easy to cause misjudgment by relying solely on the co-rate to calculate similarity,while ignoring other aspects.(3)The user's rating behavior has habits and other factors,so rating cannot be directly equivalent to satisfaction.For Improving the above problems,the main methods are as following:(1)An item-based similarity measure method adapted to data sparsity is designed.It is more difficult to find co-rate with increasing date sparsity.In this case,the efficiency of similarity measure methods based on co-rate decrease significantly.In view of this issue,the Increment of Diversity in the field of bio-informatics was imported,and a coefficient based on it was constructed.Then a method,which uses the frequency distribution of all ratings of the item,for calculating item-based similarity was proposed.It breaks the restrict of co-rate and alleviates the impact of data sparsity.(2)A multi-dimensional item-based similarity measure model is constructed.Thetypical CF methods singly use ratings to calculate the similarity of items.The comprehensiveness of the metric is insufficient.It is easy to mistakenly judge items with similar ratings while they are quite different in nature.To improve accuracy,the improved Jensen-Shannon(JS)divergence was combined to incorporate the absolute quantity factor into similarity measure.Then,the similarity in the attribute was calculated according to tags.Finally,the multi-dimensional similarity model of the fusion the density,the absolute quantity of ratings and item's attributes was proposed.(3)A preference model reflecting the real satisfaction of users is constructed.The issue is he subjective rating value cannot be directly equivalent to their actual satisfaction.So,the rule that rating value is converted into preference based on Borda Count theory,was proposed.Then,the rule was amended by adding user rating habits,preferences and other factors,and the final preference model was obtained.Experiments show that some existing recommendation algorithms have smaller error when using the transformed data of preference model.(4)A hybrid recommendation method is proposed.The users' nearest neighbor sets were generated using clustering algorithm,which with preference data as input.And,the nearest neighbor sets of items were generated by combining multi-dimensional item-based similarity measure method.Finally,a hybrid predicting and recommendation scheme using two neighbor sets was designed.In summary,the proposed method enhances the ability to resist data sparsity,improves the quality of recommendations,and may have greater application potential. |
