Research On Collaborative Filtering Algorithm Based On User Preferences And Characteristics Attributes

Collaborative Filtering is one of the most popular and most successful recommendation techniques used in recommendation systems.But it had been affected by data sparsity,cold start and other issues.In response to these problems,the traditional collaborative filtering algorithm is studied and improved in this thesis.(1)Sparse data in reality will result in lack of common assessment items between users,and cause some traditional similarity measures cannot be calculated;in addition,traditional collaborative filtering algorithm ignores the problem of user preference which leads to decrease in recommendation accuracy.To deal with these problems,this article analyzes the factors affecting user interest preferences from the user's global items and local rating information,Calculate the global probability distribution of user rating information and use Heming Post progress to calculate user interest preference,Then use Jeffries-Matusita distance to derive a similarity algorithm about user preferences.Finally,the similarity algorithm is effectively combined with traditional algorithm which propose a collaborative filtering algorithm model based on user preferences in sparse data.The experimental results show that the proposed model performance better than traditional collaborative filtering algorithms.This model has a good result even on more sparse data.(2)Under sparse data,the similarity calculation distortion problem is caused by the lack of common evaluation items.To solve this problem,the idea of confidence level is introduced: 1.Compare the user 's rating items,then Grasp the user's global preference for the project;2.Use the information entropy principle to calculate the information elements hidden in the data by the user;then,For the cold start problem,this thesis consider the implicit semantic information that affects the recommendation performance,namely the user's multi-Characteristic attributes which including the user's gender,occupation,age,and so on.The user's Characteristic attributes are specifically divided and weighted,and the distance formula is used to calculate the characteristics similarity between different users.Secondly,considering the influence of the time span on the recommendation,the logistic function is used to map the time factor and punish the scoring of the long-term items;finally,combined with the traditional algorithm,an improved algorithm for similarity of multi-user characteristics attributes is proposed.Compared with traditional algorithms on multiple data sets,the experimental results show that the improved algorithm has a lower average error.