Analytics to Combat Fraudulent Behaviour on Online Platforms

亚马逊、优步和Spotify等数字服务已经改变了我们消费商品和服务的方式。在幕后，这些平台依靠数据驱动的政策来应对各种运营挑战，例如供需匹配、定价等。然而，这些算法所做的选择可能会对平台上服务提供商的收入产生不利影响，这些服务提供商随后会被驱使操纵数据，以“玩弄系统”。例如：1)亚马逊上的卖家求助于欺诈性评论和点击机器人来提升自己的地位；2)Spotify上的艺人雇佣虚假用户来播放他们的音乐，因为他们的收入份额与他们的流媒体流量成正比。数据操纵的兴起可能会破坏人们对数字经济的长期信任，因为它可能会给数百万消费者和服务提供商带来不公平的结果。认识到这一点，学术研究人员和平台本身都在欺诈检测方面投入了相当多的资源，例如，2019年，亚马逊花费了5亿美元来打击其平台的滥用。不幸的是，这些努力往往是临时的，而且是“被动的，而不是主动的”。这项提议的重点是为在线平台开发指令性的、数据驱动的分析，以应对数据操纵。与以欺诈检测为中心的现有方法不同，该项目通过设计即使在数据损坏时也能学习最佳结果的解决方案，将健壮性放在首位。特别是，我们采取了三管齐下的方法，融合了理论和经验方法：1.引入一个新的框架来对平台上的敌对用户进行建模，并量化这些用户的经济影响。2.设计机器学习算法，该算法对处理与消费者选择和收入管理相关的问题具有很强的鲁棒性。3.介绍与电子商务、流媒体等有关的案例研究，以评估该办法并得出政策影响。最后，将针对最先进的欺诈检测算法对所提出的方法进行评估，以表征主动方法是否优于被动方法。影响：仅在加拿大，2019年数字经济就为整体GDP贡献了1180亿美元。在这方面，拟议项目的完成将标志着朝着使数字经济更可靠和防止不公平结果迈出重要一步。同时，我们的研究还将为政策制定者及时了解网络欺诈的社会影响，并指导公司如何解决这一问题。最后，该项目产生的跨学科出版物将促进最先进的技术，提供在不同级别培训HQP的机会，并将包括经济学、市场营销学和计算机科学在内的众多运营领域汇聚在一起。

Digital services such as Amazon, Uber, and Spotify have transformed how we consume goods and services. Behind the scenes, these platforms rely on data-driven policies to tackle various operational challenges, e.g., matching demand and supply, pricing, etc. However, the choices made by these algorithms can adversely impact the revenue of the service providers on the platform, who are then driven to manipulate the data in order to "game the system". For instance: 1) sellers on Amazon resort to fraudulent reviews and click bots to boost their own standing, 2) artists on Spotify hire fake users to stream their music as their revenue-share is proportional to their streaming volume. The rise of data manipulation threatens to undermine long-term trust in the digital economy as it can result in less-than-equitable outcomes for millions of consumers and service providers. Recognizing this, both academic researchers and the platforms themselves have devoted considerable resources to fraud detection, e.g., in 2019, Amazon spent $500 million to combat the abuse of its platform. Unfortunately, these efforts tend to be ad-hoc and "reactive rather than proactive". This proposal focuses on developing prescriptive, data-driven analytics for online platforms in the face of data manipulation. In a departure from existing approaches that center around fraud detection, this project prioritizes robustness by designing solutions that learn the optimal outcome even when the data is corrupted. In particular, we adopt a three-pronged approach that blends theoretical and empirical methodologies: 1. Introduce a new framework for modelling adversarial users on platforms and quantify the economic impact of such users. 2. Design machine learning algorithms that are robust to manipulation for problems relating to consumer choice and revenue management. 3. Present case studies pertaining to e-commerce, streaming, etc. to evaluate the approach and derive policy implications. Finally, the proposed methods will be evaluated against state-of-the-art algorithms for fraud detection to characterize whether a proactive approach outperforms reactive ones.   Impact: In Canada alone, the digital economy contributed $118 billion to the overall GDP in 2019. In this context, the completion of the proposed project would mark a significant step towards making the digital economy more reliable and preventing unfair outcomes. At the same time, our research will also provide policy makers with a timely understanding of the societal implications of online fraud, and guidelines on how to direct companies to address this problem. Finally, the interdisciplinary publications resulting from this project would advance the-state-of-the-art, provide opportunities for training HQP at different levels, and bring together numerous domains beyond Operations including Economics, Marketing and Computer Science.