I-Corps: Tree-based artificial intelligence (AI) models for financial fraud detection

这个i-Corps项目的更广泛的影响/商业潜力是潜在地开发用于银行和信用卡公司等金融服务的欺诈检测软件。欺诈检测是一个代价高昂的问题，每年会给银行、商家和客户造成数十亿美元的损失。另一个成本是声誉，因为在交易下降后，持卡人减少了对商家的光顾。该团队寻求为商业银行提供精确度超过行业标准的欺诈检测人工智能(AI)软件，可能为银行节省数百万美元。这项技术是基于学习决策树和森林的一项根本性创新，这使得有可能在几个方面提高欺诈检测的技术水平：卓越的预测准确性和更少的误报；更快的预测时间，每秒处理许多交易所必需的；以及可解释和可信的模型。这个I-Corps项目基于欺诈检测，可以被框架为二进制分类问题，其中交易被分类为合法或欺诈性的。目前实用的欺诈检测系统通常使用在过去交易的大数据集上训练的决策树和森林。然而，从数据中学习一棵树需要解决一个非常困难的数学优化问题。直到最近，这还是用20世纪80年代的启发式算法来近似的，这些启发式算法提供了次优模型。提出的算法基于现代优化原理，寻求在扩展到大型数据集的同时找到更好的解决方案。这项技术可能会产生更准确的模型，但也会产生更浅的树，更容易解释和审计，也可以更快地计算预测。该技术还可以学习更多基于树的模型的一般类型，这可能会导致信用评分等应用程序以及法律、政府和公共卫生部门的性能进一步提高。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

The broader impact/commercial potential of this I-Corps project is the potential development of fraud detection software for use in financial services, such as banks and credit card companies. Fraud detection is a costly problem with annual losses of billions of dollars to banks, merchants, and customers. Another cost is reputational, as cardholders reduce patronage to merchants after transaction declines. This team seeks to provide fraud detection artificial intelligence (AI) software for commercial banks with an accuracy that exceeds industry standards, potentially saving banks millions of dollars. The technology is based on a fundamental innovation in learning decision trees and forests which makes it possible to improve the state of the art in fraud detection along several fronts: a superior predictive accuracy and fewer false positives; a faster prediction time, necessary for processing many transactions per second; and models that are explainable and trustworthy.This I-Corps project is based fraud detection that can be framed as a binary classification problem, where a transaction is classified as legitimate or fraudulent. Current practical fraud detection systems often use decision trees and forests trained on large datasets of past transactions. However, learning a tree from data requires solving a very difficult mathematical optimization problem. Until recently, this was approximated using heuristic algorithms dating from the 1980s which provide suboptimal models. The proposed algorithm, based on modern optimization principles, seeks to find much better solutions while scaling to large datasets. The technology may result in more accurate models, but also in shallower trees, which are easier to explain and audit and which can compute predictions faster. The technology can also learn more general types of tree-based models, which may lead to further improvements in performance in applications such as credit scoring, and in the legal, government and public health sectors.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.