一种检测兴趣漂移的图结构推荐系统

协同过滤是构造推荐系统最有效的方法之一.其中,基于图结构推荐方法成为近来协同过滤的研究热点.基于图结构的方法视用户和项为图的结点,并利用图理论去计算用户和项之间的相似度.尽管人们对图结构推荐系统开展了很多的研究和应用,然而这些研究都认为用户的兴趣是保持不变的,所以不能够根据用户兴趣的相关变化做出合理推荐.本文提出一种新的可以检测用户兴趣漂移的图结构推荐系统.首先,设计了一个新的兴趣漂移检测方法,它可以有效地检测出用户兴趣在何时发生了哪种变化.其次,根据用户的兴趣序列,对评分项进行加权并构造用户特征向量.最后,整合二部投影与随机游走进行项推荐.在标准数据集MovieLens上的测试表明算法优于两个图结构推荐方法和一个评分时间加权的协同过滤方法.

Novel Graph-based Recommender System with Interest Drift Detection

Collaborative Filtering(CF) is regarded as one of the most successful approaches for building recommender systems.Among CFs,the study of graph-based recommendation methods has become a research hot spot.Graph-based methods consider users and items as vertices of a graph and leverage graphical theories to characterize the similarities between users and items.The recommendation list is built according to the similarities of candidate items with a given user.Though graph-based recommender systems have been widely studiedvand applied,all of them neglect an important fact that users′ interests usually change from time to time.Thus,they fail to model the dynamic interests of users reflected through user ratings.In this paper,we propose a novel graph-based recommender system with user interest drift detection.Firstly,we design a novel interest drift detection method that takes both the content variety of rated items and the change of a given user′s neighbors into consideration.This method is effective for capturing when and how a user′s interest changes.Secondly,rated items are weighted for constructing users′ feature vectors according to their interest drift series.Finally,we combine the bipartite graph projection and the random work approach for recommendation.The experiments on MovieLens data demonstrate that our recommender method outperforms two other graph-based recommendation approaches and a time-weighted memory-based collaborative filtering method.