基于信任影响力的社会推荐方法研究

The rapid proliferation of social networks, such as Facebook and Twitter, enriches people’s social activities with others, and produces potential trust relationships between people. It is reported that 84% consumers’ purchase behaviors were influenced by the recommendation from friends and family. Thus, trust relationships in social networks provide a new way to recommendation. Although social recommendation has been extensively studied, the real practice in industry does not present successful performance. One of the most possible reasons lie in that current recommendation is based on a single social network, and thus the trust relationships are usually sparse and with single semantics. In addition, current recommendation usually only distinguishes the influence from different trust relationships, but ignores the other complex characteristics of social influence. In this project, we aim at studying how to effectively leverage trust influence for recommendations. In particular, we focus on capturing the multi-typed semantics of trust relationships and multi-typed characteristics of trust influence when doing recommendation. To achieve the goal, we first integrate user information distributed in heterogeneous social networks. We solve the problem by proposing a probabilistic factor graphical model to incorporate the features extracted from dynamic user behavior sequences and directed networks. Based on the integrated network, we secondly mine trust relationships with multi-typed semantics. We propose an approximate probabilistic factor graphical model to incorporate multi-typed trust correlation features to mine the relationships efficiently. Then we design a sampling-based streaming algorithm to learn structural influence between users. Finally, we build a RNN model to do social recommendation, where attention mechanisms are leveraged to describe dynamic influence, topical influence and structural influence between users. The project aims at improving the effectiveness and efficiency of trust influence-based social recommendation.

社会网络的蓬勃发展极大丰富了人与人之间的社会关系。基于信任关系产生的影响力进行社会推荐引起了极大的关注。然而，已有的社会推荐方法仍然不能很好地应用于工业界，其原因在于一方面，已有推荐一般基于单一网络中语义单一且稀疏的信任关系，没有充分利用蕴含在多源异构网络中的复杂关系；另一方面，仅简单区分了信任关系的影响力强弱，忽略了影响力的复杂特性。本项目深入研究基于信任影响力的社会推荐，着眼于解决信任关系挖掘面临的多语义性问题与社会推荐面临的信任影响力多重特性问题。首先，构建概率因子图模型融合用户动态行为特征与网络有向传递特征，对异构网络进行融合；在此基础上，构建近似概率因子图模型融合异构网络中多语义信任关系关联特征，对信任关系进行高效挖掘；最后，设计适应流数据的图采样算法快速度量用户间的结构影响力，并构建循环神经网络模型进行社会推荐，在模型中引入注意力机制来刻画信任影响力的动态性、话题性与结构性。

本项目围绕信任关系挖掘面临的多语义性问题与社会推荐面临的噪声和缺标签问题，重点研究了融入网络关系语义的多源异构网络大数据融合方法、融入多语义信任关系传递性的高效信任关系挖掘方法以及去噪与缺标签条件下的社会推荐方法。通过项目的执行，课题组完成了预定的目标，并取得了以下三方面的成果：（1）研究了基于网络中节点与关系语义对异构社交图谱和跨语言知识图谱进行节点融合的方法，显著提升了异构社交图谱融合和跨语言知识图谱融合的效果；（2）研究了融合目标关系自身属性特征与邻居关系的多语义关联特征对网络关系信任度进行估计的方法，显著提升了大规模社交图谱中信任关系挖掘的精度和效率；（3）本项目研究了用户历史噪声行为感知的社会推荐方法以及缺标签场景下的社会推荐方法，在两种情境下都显著提升了推荐的精度。课题组在理论上取得了突破，发表了一系列高水平的论文。