Uncovering financial statement fraud detection with multi-modal positive-unlabeled contrastive learning

Abstract

Current research in AI-based financial statement fraud detection (FSFD) predominantly draws on financial and textual modalities and employs supervised machine learning. However, few attend to the inconsistency between these modalities, which can provide a critical indicator in FSFD. Moreover, the assumption underlying supervised learning is often violated, as non-fraudulent at present may later be revealed to exhibit fraudulent behavior. Accordingly, we introduce multi-modal inconsistency learning combined with Positive-Unlabeled (PU) learning, which trains a classifier using only positive and unlabeled data. Nevertheless, two challenges persist: the semantic gap between modalities in inconsistency learning and selection bias when identifying reliable negative instances in PU learning. To overcome these challenges, we propose PUMLCL: Positive-Unlabeled learning with Multi-Level Contrastive Learning. First, to bridge the semantic gap between modalities, we employ instance-level contrastive learning (CL) and introduce an inconsistency-enhancing module for the multi-modal inconsistency learning. Simultaneously, we introduce industry-level contrastive learning to learn the representations that capture distinctive industrial characteristics. Next, utilizing the learned industry-aware representations, we mitigate selection bias by identifying reliable negative instances with respect to the positive ones across industries. Finally, we design a classifier augmented with group-level contrastive loss to enhance the separability between fraudulent and truthful companies. Extensive experiments on real-world datasets validate the effectiveness of PUMLCL.