异质行为数据融合的电子商务虚假评论者识别研究

As the rapid growth of e-commerce, review spammers who post fake reviews to influence the purchases decision to gain economic profits became increasingly rampant, which seriously disrupting the order of e-commerce. Review spammer detection, therefore, has become a hot topic in both academia and industry. Most of the existing works have defined a set of effective behavior features, upon which various classifiers could be constructed to identify spammers. However, spammers are becoming more and more cunning, and thus the behavior of them is almost the same as that of normal users. So, the detector merely based on distinctive features is probably to be failed. This declared project aims to incorporate the relational network and review content into the traditional behavior features, and thus explore spammer detection methods by fusing these heterogeneous behavior data. In particular, we will first extract and model the heterogeneous user behavior data, including behavior features, relational network and review content. Then, we will construct a probabilistic description model and objective function to fuse the heterogeneous behavior data, and design detection algorithm based on the partially supervised learning. Finally, a case study on real-life e-commerce platform will be applied, where we will also try to optimize the detection results and management strategies to take into account the profits of the platform, which will motivate the review spammer detection and provide the decision support for the reasonable spammer control. This declared project would be expected to complement and promote both theoretical research and practical applications of the review spammer detection.

随着电子商务的发展，利用虚假评论影响用户购买决策以攫取经济利益的虚假评论者日益活跃，严重扰乱了电子商务运营秩序，虚假评论者识别因此成为当今学术界和工业界广泛关注的研究热点。国内外已有研究多利用行为特征构建分类器判定评论者类别，但随着攻击方式的演进，虚假评论者的行为特征不断趋近于正常用户，单纯依靠行为特征难以识别这些隐蔽虚假评论者。本项目将进一步融入关系网络和评论内容，研究异质行为数据融合的检测方法，提升隐蔽虚假评论者的识别能力。首先，分别对行为特征、关系网络、评论内容等异质行为数据进行挖掘与建模；其次，构建融合上述数据的概率模型与目标函数，设计基于部分监督学习的检测算法；最后，在真实电商平台中进行应用案例研究，探索兼顾平台收益的检测结果及管理策略优化方法，提升平台检测虚假评论者的动力并为进一步的合理管控提供决策支持。本项目的研究有望对虚假评论者识别理论和应用实践提供重要的补充和推动作用。

随着电子商务的发展，利用虚假评论影响用户购买决策以攫取经济利益的虚假评论者日益活跃，严重扰乱了电子商务运营秩序，随着攻击方式的演进，虚假评论者的行为特征不断趋近于正常用户，单纯依靠行为特征难以识别这些隐蔽虚假评论者。本项目融合关系网络等异质行为数据，研究基于多视图数据混合学习的检测方法，有效提升了隐蔽虚假评论者的识别能力，并在大数据计算性能及应用场景两方面对混合学习方法进行了提升和扩展，主要成果内容包括：（1）异质行为数据的分析建模与网络关系挖掘，开展多关系社会网络恶意用户行为分析，建立行为序列模型，提出基于泛化指标及局部学习的社区检测框架及基于节点影响力的动态网络社区发现方法，有效构建并优化了关系网络；（2）异质数据融合学习模型与检测算法，提出行为特征与关系网络混合学习的虚假评论发布者检测框架，基于余弦模式挖掘的水军群组检测方法，及基于流量关联的移动应用识别及恶意软件检测方法；（3）面向海量数据的混合学习方法实例化应用与扩展，在海量数据处理方面，提出基于关联规则推荐的高效分布式计算框架，面向高速数据流分析的自适应Sketch框架，在应用扩展方面，提出融合异质数据的旅游产品及行程推荐方法。本项目的研究在隐蔽虚假评论者检测，及关系网络与行为特征的融合学习上获得了一定的理论突破，从而形成了以下几个方面的成果：（1）发表期刊论文16篇，会议论文4篇，SCI收录12篇，包含6篇IEEE/ACM汇刊论文；（2）获得授权3项国家发明专利；（3）培养研究生6名。