Deep Latent Representation and State Recursion for Enterprise Audit Risk Detection

Abstract

In highly complex and data-intensive corporate operating environments, audit risk increasingly exhibits hidden structural patterns, temporal evolution, and multi-actor interactions. Traditional audit analysis methods that rely on static indicators and manual rules struggle to capture how risk forms and accumulates over time. To address this challenge, this study proposes a dynamic enterprise audit risk detection approach based on deep representation learning. The method performs unified modeling of multi-source corporate behavior data and maps enterprise operating states into a latent representation space. Risk states are continuously updated and characterized along the temporal dimension. Deep nonlinear mappings are used to automatically extract high-level semantic features. These features describe latent structural relationships and dynamic changes in corporate transaction behaviors. A state recursion mechanism is introduced to model the evolution of audit risk in a systematic manner. This avoids information fragmentation caused by single-point judgments. In addition, a unified risk scoring mechanism is constructed to express risk states in a continuous and comparable form. This enhances discrimination consistency and stability in complex audit scenarios. Comparative experimental results show that the proposed method outperforms representative existing approaches in overall identification performance and risk differentiation capability. The findings validate the effectiveness of combining deep representations with dynamic modeling for enterprise audit risk identification. This work provides a systematic solution for shifting audit risk analysis from static assessment to dynamic perception. It also offers a valuable reference for data-driven intelligent auditing research.