Research On Collaborative Filtering Recommendation Algorithms Based On Multi-level Similarity And Information Core

With the rapid development and popularization of the Internet,the data resources is increasing exponentially,which makes users unable to select useful information for themselves efficiently when facing huge information resources,and then the information overload problem arises.So the personalized recommender systems emerges as the times require.It carries out personalized calculation by analyzing users' interests and preferences based on users' historical information,so as to provide useful information for users.Nowadays,recommender systems have been widely used in social networks,e-commerce,e-learning,film recommendation,tourism and many other fields.With the in-depth study on recommender systems,the research problems emerge,such as the cold start problem when recommending the new registered users,the data sparsity problem of user rating matrix,the system scalability problem caused by the surge in the number of users and items,and lack of diversity in most similar items recommendation,which all affect the further development of recommender systems.In this paper,Aiming at the increasingly serious scalability problem,recommendation time-consuming and accuracy of similarity function in collaborative filtering system,we have done the following research:(1)Aiming at the scalability problem of collaborative filtering recommendation,this paper uses the method of extracting information core.That is to say,we train all users in a user set to extract the most valuable core users to form an information core.This process is operated offline,which will greatly save memory and time when calculating user similarity.In this paper,we propose two improved information core extraction methods IFB(IFrequency-based)and IRB(IRank-based,IRB)based on the original FB and RB information core extraction methods,and we propose an optimization set concept when searching for the most similar neighbors.In the optimization set,we use two parameters items rates and user similarity to find the most similar neighbors for each user.This algorithm also reduces the time consume greatly at the same time in the recommendation process.(2)Aiming at the accuracy of similarity calculation function in CF,We improves the traditional measurement standard Pearson Correlation Coefficient(PCC),and proposes a multi-level similarity algorithm(MLPCC).The algorithm is divided into five layers,each layer corresponds to different constraints and adjustment parameters.On the basis of calculating user similarity by using items rates,we fully consider the number of co-rating items which effects the user similarity too,so that users with more co-rating items will have higher similarity value.So we optimize the measurement standard.The results show that compared with other recommendation algorithms,the proposed method can effectively overcome the scalability problem,save a lot of calculation time andreduce the average absolute error(MAE).Additionally,it has higher Precision and Recall,so the recommendation performance is better.