Research On Collaborative Filtering Recommender Algorithms Based On Neighborhood

Collaborative filtering recommendation is a key technology to solve the information overload problem.It mainly uses the behavior data of the existing user group to predict the objects that the current user likes most or may be interested in.Among them,as the most successful collaborative filtering recommendation algorithm,the neighborhood-based method is widely used in e-commerce websites,social networks,video websites,and has become the research focus in the field of collaborative filtering recommendation.However,the current research has the following problems:(1)Aiming at the reduction of recommendation accuracy and coverage caused by sparseness of rating data and cold start of users,the trust relationship between users in social network can be used to alleviate this problem,but the trust relationship has not been fully integrated,resulting in a lack of a perfect trust relationship recommendation model;(2)In the study of the user being the object,when selecting the nearest-neighbors set,on the one hand it does not consider the impact of the changes of target items on the choice of neighbors,on the other hand it also ignores the contributions of neighbor users recommendation ability on the target user;(3)In the study of the item being the object,it usually solves the problem of inaccuracy in single similarity calculation with the interest relationship between items,which leads to the complex quantification of the interest relationship between items.Therefore,this paper focuses on the above issues,and the main research work and results are listed as follows:1.A collaborative filtering recommendation algorithm based on multiple trust is proposed.Firstly,an improved Mean Squared Difference trust measuring method is proposed based on the accuracy and dependability factor of the recommended ratings among users.Based on this,a rating model based on implicit trust information is proposed.Secondly,with the maximum trust propagation distance as the constraint,a relation model based on explicit trust information is proposed.Finally,based on the rating similarity and the explicit trust relationship,the neighbor set of the target user is accurately selected by the 0-1 backpack combination optimization strategy,and the rating prediction is carried out.Comparisons of the simulation results with a variety of state-of-the-art algorithms on Epinions dataset demonstrate that the proposed algorithm can greatly alleviate the data sparsity and cold start problems by introducing effective rating and explicit trust relationship,and significantly improve the recommendation accuracy while preserving a good coverage.2.A collaborative filtering recommendation algorithm based on entropy optimization nearest-neighbor selection is proposed.Firstly,the algorithm uses the Bhattacharyya Coefficient to measure the similarity between items,and then calculates the similarity between users by weighting the similarity between items.Secondly,it introduces the information entropy to define the distribution characteristics of the user's rating,and measures the contribution of neighbor users' recommendation ability on the target user by the difference between the rating distributions;Finally,it combines the users similarity and the contribution of recommendation to calculate the recommendation weights,and builds the nearest-neighbor set on them.The experimental results on the MovieLens 1M dataset show that,in the study of the user being the object,the proposed algorithm can accurately select the nearest-neighbor set without sacrificing the time complexity,and significantly improve the accuracy of recommendation.3.A recommendation algorithm based on combining the number of co-rating users and interest relationship of items is proposed.Firstly,it improves the Pearson similarity method by using the similarity between items and the number of common rating users;Secondly,it effectively introduces the interest relationship between items to define the direct interest degree with the user's rating information on the item.Then it takes the users in the system as the bridge to calculate the indirect interest degree,and then proposes a similarity model based on the item interest relationship;Finally,it weighted-fits the interest relationship and similarity between items,and designes the recommendation algorithm on this basis.The experimental results show that in the study of the item being the object,the proposed algorithm can significantly improve the recommendation accuracy and can greatly alleviate the influence of data sparseness on the similarity calculation by effectively introducing the interest relationship among the items.