Graph Signal Processing and Graph Machine Learning

尽管电子商务的进步不断改善经济和人民生活，但个人和企业经常成为欺诈金融交易的受害者，估计每年损失分别为70亿GB和400亿GB。尽管开发一个高效的监管系统来发现和防止此类欺诈性交易很重要，但金融网络的复杂性和海量使得对这类数据进行适当建模变得困难。在这项研究中，我们的目标是使用图神经网络(GNNS)的机器学习技术来构建有效的金融监管人工智能模型，该模型将金融网络中的账户和交易建模为节点和链接，然后通过网络传播信息。该方法的主要特点是，当处理单个节点(即金融账户)的信息时，该模型还会收集其邻居的行为信息(即有交易的账户)，而不需要在整个网络上进行搜索，这使得该模型成为从数百万笔金融交易中发现欺诈性集群的理想方法。特别是，我们的研究将集中在表示方法上，它本质上是一个过滤器，可以筛选出金融网络中丰富的无关信息，然后使用剩余的特征作为表示来执行下游任务。除了欺诈检测，我们的研究还涉及金融网络的其他监管领域，我的研究提案中涵盖了技术细节。在信用风险管理中，我们可以将各种金融实体(例如银行、公司、个人和政府)及其关系(例如贷款、借款和违约)建模为银行贷款网络。然后，根据实体的金融历史，将它们嵌入到低维表示中，分析它们的行为，预测贷款的违约概率，帮助贷款人做出关于新贷款的决策，同时监控现有的贷款状况。在股票市场监管中，我们可以把账户和金融产品看作一个以买卖行为为纽带的二分图。然后，通过将图形编码到嵌入中，我们将能够识别高度关联或具有相似市场行为的实体集群，这可以帮助监管机构发现市场中的异常情况，如操纵、内幕交易和其他非法活动。

Despite the advancement in e-commerce continuously improving the economy and people's lives, individuals and business often become victims of fraudulent financial transactions with an estimated annual loss of £7 billion and £40 billion respectively. While it is important to develop a highly efficient regulation system to detect and prevent such fraudulent transactions, the complexity and high volumes of financial networks make it difficult to properly model this type of data. In this research, we aim to build effective AI models for financial regulation using machine learning techniques of Graph Neural Networks (GNNs), which model accounts and transactions in thefinancial network as nodes and links, then propagate information through the network. The key characteristic of this method is that when processing information of an individual node (i.e., financial account), the model will also gather information of its neighbours' behaviour (i.e., accounts with transactions) without searching over the entire network, making it an ideal approach to detect fraudulent clusters out of millions of financial transactions. In particular, our research will focus on Representation methods, which essentially work as a filter that can "screen out" the abundant irrelevant information in the financial network, and then use the remaining features as the "representations" to perform down-stream tasks. In addition to fraud detection, our research also works on other regulatory domains of financial networks, with the technical details covered in my research proposal. In credit risk management, we can model various financial entities (e.g., banks, companies, individuals and government) and their relationships (e.g., lending, borrowing and default) as a bank lending network. Then, with the entities' financial history, we can embed them into low-dimensional representations to analyse their behaviours and predict the default probability of loans, which helps lenders to make decision about new loans and monitor existing loans status at the same time. In stock market surveillance, we can treat the accounts and financial products as a bipartition graph with purchase/sell actions being the links among them. Then, by encode the graph into embeddings, we will be able to identify the clusters of entities that are highly correlated or have similar market behaviour, which can help regulators to detect anomalies in the market such as manipulation, insider trading, and other illegal activities.