Adaptive Detection of Shill Bidding and Multi-Objective Winner Determination

电子(E)拍卖中需要解决的两个重大研究挑战是：监测拍卖中的欺诈行为，以及确定高级拍卖的获胜者。由于用户的匿名性和涉及的金额巨大，电子拍卖对诈骗者非常有吸引力。发现电子拍卖中的欺诈行为具有挑战性，因为它们构成了一个庞大而充满活力的市场。当有人出价人为地提高物品的价格时，上当竞价(SB)是最难发现的拍卖欺诈，因为它与通常的竞价行为相似。将机器学习应用于检测SB的研究很有限，而且大多数都是离线进行的。随着时间的推移，自适应学习对于提高分类器的性能是必要的，这对于欺诈检测问题是至关重要的。我们的目标是设计一种解决方案，解决目前SB检测的挑战。我们将开发一个高质量的标签、采样的SB训练数据集，这是目前缺乏和不可用的；开发一个在线的、自适应的分类器，它将随着新的投标趋势不断演变；以及开发一种欺诈验证方法，因为训练数据是在没有任何基本事实的情况下使用的。我们的动机是基于最先进的技术有效地解决这些问题，这些技术包括数据集群和标记、不平衡数据采样、增量和减量分类以及自动欺诈验证。学习的SB模型根据需要对从eBay抓取的新数据进行分类，eBay每天举行数千次电子拍卖。它将在竞标期结束时但在确定中标者之前推出，以避免经济损失。我们的SB分类器将使用有信心的标记数据不断调整，这将提高其性能和训练数据的基本真实性。*组合反向拍卖(CRA)由于其配置效率，代表了一种经济的采购方法。CRA允许卖家对单一买家所需的捆绑商品或服务进行竞价。我们的目标是为高级信用评级机构开发一种稳健的赢家确定(WD)方法，目前这些机构受到交易限制和相互冲突的目标的限制。我们的问题本身就是目标最大化和最小化的混合体。我们将WD形式化为一个多目标优化(MOO)问题，我们为其寻找一组权衡解决方案；每个方案同时优化相互冲突的目标。但是，执行这种权衡分析需要大量的计算。在拍卖环境中，执行时间是一个关键要求。为了应对这一挑战，我们将定义一种基于MoO的渐进式WD方法。作为一个真实的案例，我们将为电力市场量身定做新的WD方法，以从不同的能源中最佳地获取电力，并评估其性能。此外，我们还将基于大规模的CRA实例对WD方法进行性能分析。

Two significant research challenges to be addressed in electronic (e) auctions are: monitoring auctions for fraud, and determining the winners for advanced auctions. Due to the anonymity of users and large amounts of money involved, e-auctions are very attractive to fraudsters. Detecting fraud in e-auctions is challenging because they constitute a voluminous and dynamic market. Shill Bidding (SB), which occurs when someone places a bid to artificially increase the price of an item, is the hardest auction fraud to detect due to its similarity to usual bidding behavior. There are limited studies on applying machine learning to detect SB, and most of them have been conducted offline. Adaptive learning is necessary to improve a classifier's performance over time, which is crucial for fraud detection problems. Our goal is to devise a solution that addresses the current challenges of SB detection. We will develop a high-quality labelled, sampled SB training dataset, which is currently lacking and unavailable; develop an online, adaptive classifier that will evolve continuously with new bidding trends; and develop a fraud verification method as training data are used without any ground truth. We are motivated by solving these problems efficiently based on state-of-the-art techniques for data clustering and labeling, imbalanced data sampling, incremental and decremental classification, and automated fraud verification. The learned SB model classifies on demand new data crawled from eBay where thousands of e-auctions are held daily. It is to be launched at the end of the bidding period but prior to determining the winners to avert financial loss. Our SB classifier will be constantly adjusted with confidently labeled data, and this will improve its performance and the ground truth of training data.******Combinatorial Reverse Auctions (CRAs) represent an economical procurement method due to their allocative efficiency. CRAs allow sellers to bid on a bundle of goods or services required by a single buyer. Our aim is to develop a robust Winner Determination (WD) method for advanced CRAs, currently subject to trading constraints and conflicting objectives. Inherent to our problem is a mixture of maximization and minimization of objectives. We will formalize the WD as a Multi-Objective Optimization (MOO) problem for which we search a set of trade-off solutions; each one optimizes the conflicting objectives simultaneously. However, performing this trade-off analysis is computationally intensive. In an auction setting, the execution time is a critical requirement. To address this challenge, we will define an evolutionary MOO-based WD method. As a real case study, we will tailor the new WD method to the electricity market to optimally procure power from different energy sources, and assess its performance. Moreover, we will conduct a performance analysis of the WD method based on large-scale instances of CRAs.