Enterprise Audit Risk Identification Through Temporal Context and Entity Relationship Anomaly Modeling

Abstract

This paper addresses the challenges of concealed risk clues, scarce labels, and coupling of heterogeneous information from multiple sources in enterprise auditing scenarios. It proposes an audit risk identification framework based on anomaly detection, simultaneously characterizing anomalies at the attribute, temporal context, and relational structure levels from a unified modeling perspective. First, robust preprocessing of audit records weakens scaling differences and extreme value interference. Representation learning is then used to map records to a latent space, reconstructing the deviation measure to assess the strength of deviations from normal patterns. Subsequently, a time window mechanism is introduced to construct a local contextual baseline, calculating the contextual deviation of the latent representation relative to the baseline to identify unconventional behavioral fragments within the business context. Simultaneously, a subject relationship graph is constructed, and structural information is aggregated to form a subject structural representation, which uses structural deviations to characterize abnormal associations and suspicious links. Finally, multi-source deviation signals are fused into a unified risk score, and probability mapping is used to obtain a risk intensity expression that can be used for verification ranking and threshold decision-making, thus outputting a set of high-risk records and their associated clues. Comparative experimental results show that this method outperforms other methods in terms of accuracy, precision, recall, and AUROC, validating the effectiveness and stability of the multi-source anomaly characterization and fusion scoring mechanism in audit risk identification tasks.