One-Class Classifiers Ensembles for Detecting Fund Misuse Problems within Financial Auditing

Fund misuse poses a significant challenge in fi-nancial auditing, necessitating the development of precise and efficient detection strategies. In our study, we have framed the detection of fund misuse as a multi-class classification problem, focusing on a select subset of payment data. This approach, tailored to the specifics of the domain, lays the groundwork for creating robust data analysis techniques within auditing. Our methodology begins with the use of a Large Language Model (LLM) to embed textual data, followed by the introduction of a classification framework that utilizes ensembles of one-class classifiers. Specifically, we transform the multi-class classification problem into a series of one-vs-all binary tasks, employing one-class SVM as the primary classifier. We then propose two viable one-vs-all ensemble methods: the straightforward Maximum Confidence strategy and the more complex stacking technique. For the stacking approach, we construct a three-layer fully connected neural network to serve as a meta-model, integrating the outputs of multiple base classifiers. Our experiments, conducted on two distinct real-life datasets, have validated the effectiveness of our model in pinpointing fund misuse instances. Furthermore, we explore the application of our methodology in financial auditing and demonstrate its adaptability through successful transfer learning across various datasets.