摘要
近年来随着互联网的快速发展,各种虚拟社区不断涌现,用户组成群共同活动的现象逐渐增多,人们开始逐渐关注面向群的推荐.已有的群推荐方法大多是在基于内存的协同过滤推荐方法上进行改进,或是认为组内成员相互独立,忽略了群内成员间的关联关系对群推荐结果的影响.为此,该文提出了一种基于联合概率矩阵分解的群推荐方法,更好地对群推荐问题进行建模.首先,利用用户加入的群的信息计算用户之间的相关性,其次,将用户相关性矩阵融入到概率矩阵分解过程中,得到个人预测评分,最后,利用面向群推荐问题中常用的合成策略对个人预测评分进行融合,得到群对项目的预测评分.进一步将该文提出的方法与现有常用的群推荐方法进行比较,在CiteULike数据集上进行实验,实验结果表明,该文所提出的方法在准确率、召回率等多种评价指标上都取得了更好的推荐结果.
In recent years,a lot of virtual communities are emerging with the rapid development of the Internet.However,with the ever-increasing number of the users and generated information,there is difficult for users to find the valuable interesting information on the Internet.Recommendation system has become one of the most important tools to overcome these information overload problems.Meanwhile,the users on the virtual communities gradually intend to establish a group or join certain like-minded groups to facilitate their communication and sharing,which makes the group-oriented recommendation being hotter topics in these days.Researchers begin to pay more attentions to the group recommendation system.On the one hand,existing group recommendation methods are mostly improved by memory-based Collaborative Filtering(CF)method,but the memory-based CF method is seriously affected by the data sparse problem.On the other hand,the interactions among members in a group have not been effectively utilized in existing group recommendation methods,they ignore the influence of the relationships among group members since they just considered the group members were independent of each other.Actually,users in the same group should not be independent but have certain similarities in their preferences.To solve these problems,the model-based CF method,i.e.,Probabilistic Matrix Factorization(PMF),is utilized to alleviate the data sparsity problem through adding side information into the model when predicting individual members’preferences.And the group information including users’common group number and common group size are considered when measuring the users’interactions.Therefore,a novel group recommendation method based on PMF is proposed and presented as a prettier way to model the group recommendation problem in this paper.Firstly,the users’correlations are obtained by incorporating the group information into the measurement.Having assumed that the more common groups the users have and the smaller size of the common group,the higher similarity of users.Secondly,the users’correlations which contain the group information are incorporated into the PMF model to get a better individual prediction value.Finally,to aggregate the individual prediction values into the recommendation list for the group,the mostly used group aggregation strategies,such as the Average strategy,the Least Misery strategy,and the Most Pleasure strategy,are utilized in the aggregation phase.All items’aggregation value for a whole group are sorted in the descending order,and the top N items are selected and recommended to the group.To evaluate the effectiveness and the feasibility of the proposed method,the experiments were conducted on the CiteULike dataset.Specially,the results were evaluated in terms of Precision and Recall,together with two rank-sensitive metrics,i.e.,Mean Average Precision(MAP),and Mean Reciprocal Rank(MRR).After that,several parameters were discussed including recommendation number,latent factor dimension,and regularization.The experimental results show that the proposed method in this paper has achieved better results at the evaluation metrics including Precision,Recall,MAP,and MRR.It is indicated that the proposed method which considering both the PMF model and the group information can efficiently improve recommendation performance.
作者
王刚
蒋军
王含茹
杨善林
WANG Gang;JIANG Jun;WANG Han-Ru;YANG Shan-Lin(School of Management, Hefei University of Technology, Hefei 230009)
出处
《计算机学报》
EI
CSCD
北大核心
2019年第1期98-110,共13页
Chinese Journal of Computers
基金
国家自然科学基金(71471054
91646111)
安徽省自然科学基金(1608085MG150)资助~~
关键词
群推荐
用户相关性
群组信息
概率矩阵分解
合成策略
group recommendation
user correlation
group information
probabilistic matrix factorization
aggregation strategy
