数据驱动的电商信誉评分机制及群组虚假评分检测模型研究

With the rapid development of E-commerce, a huge number of commercial deals highly depend on the reliable online reputation rating systems which assign reasonable and valuable rating scores. However, there are still challenging problems of the existing reputation rating systems. On one hand, a large number of ratings and comments are randomly given by careless consumers. On the other hand, spamming group attacks giving extremely low or high ratings usually extremely mislead the decision-makings of potential consumers. Thus efficient and reliable reputations systems which can give reasonable ratings to both users and commodities, under the challenges of random ratings and spamming group attacks, have deep theoretical insights and significant social and economic values. In this project, we aim at building the reputation system using the model of bipartite networks containing users and items initially. Then with the analysis of bimodal and co-bursting rating behaviors and their probability distributions, we can design a robust rating algorithm. The model can deal with both static rating network data sets and dynamic rating network data sets, using an iterative way to give out reasonable ratings to both users and items. Finally, real data sets will feed to the case studies. We will further explore the feasible tactics to balance the profits among the platform, trade companies and consumers. Those tactics and solutions will greatly support the justice and fair of the market. The study of this project has valuable meanings to regulate the market order and establish sound operation mechanisms of the E-commerce eco-systems.

随着电子商务的迅速发展，大量的商品交易依赖于可靠的信誉系统对物品给出合理有价值的评分。而当前的信誉评分系统面临着诸多挑战，比如用户评分的随意性引发的不合理问题，特别是有组织的恶意群组对特定商家故意提高或降低分值，严重误导消费者决策。如何建立一个高效可靠的用户和商品信誉系统，既能够识别那些有组织的恶意攻击用户，又能够对商品给出合理的评分，具有深刻的理论意义和重大的社会经济价值。本项目首先从复杂网络的二部图出发对信誉系统的两个主要元素，用户和商品进行建模。然后通过对恶意群体攻击的双峰性和并发性进行分析，设计基于关键节点识别的评分机制和时间序列模型，研究信誉系统的动态性。最后使用电商平台的真实数据进行案例分析和研究，探索如何平衡平台，商家和用户三方利益的管理策略，为电商市场的公平，公正及合理性提供有力支持。本项目对规范电商市场秩序，建立良性的电商生态有重要意义。

在电子商务的蓬勃发展中，评分系统对于商家信誉的公正排序起到非常重要的作用。一小部分水军的在非法利益的驱使下，通过违规操纵评分，影响了市场的公平和公正。.本项目基于真实的电子商务评分数据集，比如Amazon和Yelp等，并结合国内的数据集，比如豆瓣等。主要标志性成果有以下3个。(1) 发现了一种新的水军评分的统计行为模式，并提出了一种基于该模式的电商评分新算法。(2) 发现了商品评分的稀疏性，极大地影响了水军的发现。水军通过评分较少的商品进行伪装，从而操纵商品的评分。提出了基于Peason相关性和深度置信神经网络的2种新方法。(3) 结合图卷积神经网络，发现基于领域的精准化采样策略，能够较高精度检测到水军。综上所述，本项目结合网络科学二部图中水军的异常的统计行为特征，并且研究了图神经网络在电商水军中的新方法。该项目对于电子商务水军的精准检测有重要的实际意义。本项目总共发表论文8篇，其中SCI4篇，EI4篇，后期在投论文3篇。申请专利并获得授权3项。