面向群组的社交follow推荐方法研究

Group Recommendation Approach for Social Follow Relationship

目前社交网络中的推荐方法主要是针对单个的个体用户,然而随着日益频繁的社交活动,若干相关用户自然形成了群组,研究如何对于整体的群组用户进行推荐的问题引起了国内外学者的兴趣.本文在优化和改进面向个体的社交follow关系推荐算法的基础上,提出基于矩阵分解和混合策略的群组用户推荐方法.首先,对社交媒体微博中用户-项目矩阵进行奇异值(SVD)分解,然后提出在SVD中加入用户的社交行为和社交关注关系特征,并用随机梯度下降(SGD)算法优化对个体用户推荐的预测评分.其次,在获得个体的推荐评分基础之上,设计一种混合融合策略,该策略融合群组中个体成员的推荐评分形成对该群组的整体评分,从而实现对于群组用户的推荐.最后,实验采用KDDCUP2012竞赛Track1的数据,以平方根误差为评估指标,对比个体推荐中传统SVD模型和本文提出的SVD优化模型,并进一步对比本文提出的混合融合策略与传统的最大满意度、平均满意度及最小忍耐度三种单一融合策略.实验结果表明SVD优化模型优于传统SVD模型,并且采用混合的策略要优于单一的群组融合策略.总体上,本文提出的推荐方法能够有效提高群组推荐的准确度.

Currently,recommending for individual users is the main research topic in the social recommendation area. Increasingly frequent social activities make users form social groups. However the study for group recommendation is relatively few and is attracting the interests of researchers in recommender system field. In this paper, we improve the traditional individual recommendation algorithm for recommending social follow relationship, and then put forward a group recommendation approach based matrix decomposition and hybrid aggregation strategy. Firstly,based on SVD model, we analyze the characteristics of these micro-blog users, and add the social behavior and the social relationship of a user into the model. Then SGD algorithm is used to learn the optimal model parameters for individual recommendation. Secondly, for those user groups in social media, we use fusion prediction method based on the recommendation results of our improved SVD algorithm to give recommendations for group users. We aggregate the prediction rating of each member into a group rating by the most pleasure, the average strategy, the least misery and our hybrid aggregation strategy. Our experiments are conducted on the Track 1 of KDDCUP 2012 public data set. Experimental results from the different aggregation strategies show that adding the social behavior and the social relationship to SVD can improve the quality of individual recommendation, and the hybrid aggregation strategy outperforms other single strategies for group recommendation.