Research On Collaborative Filtering Recommendation Algorithm Based On KL Divergence

With the popularity and deep application of the Internet and mobile Internet,the amount of information is increasing sharply.How to solve the information overload and meeting the personalized needs of users has become one of the research hotpots.Recommendation algorithm as the core of recommendation systems(RS)has been widely applied and researched.Collaborative filtering(CF)technique is one of the most successful recommendation algorithms.In the recommendation algorithm,the most important step is to calculate the similarity between the users or items,which directly determines the quality of recommendation systems.However,most of similarity measures have the following problems when computing similarity values.(i)These methods only consider the co-rated items between two users,and ignores the rest of ratings that propbably hide valuable information.(ii)There is no linear relationship between variables,and linear models cannot be used to calculate the similarity.(iii)The user preferences and the number of ratings are absolutely different,but most of similarities cannot distinguish the different preferences between users.For improving the above key issues in CF,the main contributions of this paper are as follows:First,research on the item-based similarity measure.The traditional item-based similarity methods must depend on the co-rated items when computing similarity values between two items,which leads to low predition accuracy in sparse datasets.In view of this problem,we introduce the Kullback-Leibler(KL)divergence in the signal processing field to the item similarity,and calcuate the similarity between items from the point of probability density distribution.The experimental results show that our method uses all ratings information effectively to address the sparse dataset issue.Thus,it has strong potentiality in practice.Second,reseach on user similarity collaborative filtering algorithm based on KL divergence.User similarity based collaborative filtering algorithm is one of most widely used technologies.However,most of user similarity algorithms only consider the corated items between two users,and ignore other ratings that probably hide valuable information.To take full use of all user ratings,we propose a user similarity collaborative filtering algorithm based on KL divergence.Our scheme not only utilizes the co-rated items,but aslo uses other no co-rated items to calculate similarity between two users.The experimental results show that the proposed method has better flexibility,and improves the recommendation quality.Third,research on a hybrid similarity model based on KL divergence.The traditional similarity methods only consider the ratings,and ignores the relationship between items.Moreover,most of similarities also do not consider the influences of the number of user ratings and user preference behaviors.For solving these issues,we propose a hybrid similarity model for CF that combines an improved non-linear model PSS and item similarity based on KL divergence,and adds two weighting factors to distinguish the rating preference between users effecrively.Our model is asymmetric,and breaks the constraints of depending on the co-rated items.The experimental results show that our proposed scheme can improve the effectiveness of recommendation systems.