Collaborative Filtering Recommendation Algorithm Based On Dynamic Clustering Of User Interest Drift

Among most of the recommender systems, User-based collaborative filtering is an importantpersonalized recommendation system, which is considered that the recommendation of the targetuser’s nearest neighbour are easy to be accepted. Therefore, how to find the target user’s nearestneighbour is a key problem in User-based collaborative filtering recommender system. So far, themethod to find the target user’s nearest neighbour is clustering users based on user’s history score.Because the user’s history score can reflect the user’s interest clustering the users according to theuser’s history score can bring together users with similar points of interest and then all users in thesame cluster mutually interact with nearest neighbour. However, if at some point before theclustering some of the user’s interest shifts, it may get an inaccurate result according to the wholehistory of user rating clustering. In order to solve this problem, some scholars have proposed inrecent time windows clustering in which user’s rating only use the most recent data within the timewindow. Although the method can locate the user’s latest point of interest, data sparsely problemwill become more serious and ultimately affect the accuracy of clustering. This paper proposes auser-clustering algorithm against user interest-offset which may lead to inaccurate of cluster. Themain contributions of this paper include：（1）Propose an item characteristic network diagram. Item-cluster as the basic unit in the itemmanagement. Item-cluster is established by mining the similarity of inherent and hidden featuresamong all items, which meets the intra-cluster high cohesion and low coupling characteristicsbetween clusters. And use the item-cluster manage items is more efficient than using item-node.（2）Identify the multi-interest offset users. To find the variation of user’s interest in thedynamic process and then identify the user’s current real interest, thus it can avoid the impact ofcluster caused by the user’s interest-offset.（3）Alleviate the passive migration of user interest. According to interest shift users, modifythe user score with satisfaction, eventually, using the modified user rating to collaborative filtering.（4）Alleviate the data sparsely. Special processing of user rating data is only that users whoseinterest shifted, other historical data directly involved in user clustering, and not just use the data inthe current time window.（5）Compare this algorithm with others respectively in the aspect of the accuracy of clusteringand recommendation. The experiment proved that the algorithm in this paper has better clustering efficiency and recommendation results in the same condition.