Unsupervised Graph Modeling for Anomaly Detection in Accounting Subject Relationships

2026-04-30 · Source: cs.LG updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Financial Accounting & Auditing · Depth: Expert, quick

Summary

A new unsupervised discriminant framework, based on graph neural networks, has been proposed for anomaly detection in accounting subject association structures. This framework models accounting subjects as graph nodes and their co-occurrence and debit/credit correspondence in business records as weighted edges, forming a period-level accounting subject association graph. Edge weights are characterized by statistical measures like co-occurrence frequency or amount aggregation. The method uses a message passing mechanism for representation learning to obtain node embeddings with structural information. For anomaly detection, a relation reconstruction decoder estimates the rationality of subject pair connections, defining edge-level anomaly scores based on reconstruction probability deviations. These scores are aggregated to provide node-level risk rankings and localize anomalies, capturing both local substructure anomalies and cross-community connections without requiring anomaly labels. Comparative experiments show stable discriminant capabilities and high top-ranking accuracy.

Key takeaway

For forensic accountants or auditors analyzing complex financial data, this unsupervised graph modeling framework offers a robust method to identify subtle anomalies in accounting subject relationships. You can use this approach to pinpoint structural deviations and suspicious connections without relying on pre-labeled anomaly data, significantly enhancing the efficiency and accuracy of fraud detection and risk assessment.

Key insights

Graph neural networks can effectively detect unsupervised anomalies in accounting subject relationships by modeling structural deviations.

Principles

• Accounting subjects form weighted graphs.
• Message passing fuses node attributes and context.
• Reconstruction deviation indicates anomaly scores.

Method

Abstract accounting subjects as graph nodes and co-occurrences as weighted edges. Use message passing for node embeddings. Employ a relation reconstruction decoder to estimate connection rationality and define edge-level anomaly scores, then aggregate for node-level risk.

In practice

• Identify structural deviations in ledger details.
• Detect cross-community anomaly connections.
• Output traceable subject pair risk clues.

Topics

• Anomaly Detection
• Accounting Subject Relationships
• Graph Neural Networks
• Unsupervised Learning
• Graph Modeling

Best for: AI Scientist, Research Scientist, Machine Learning Engineer

Related on AIssential

• Towards Anomaly Detection on Relational Data — RelAD offers a reconstruction-based framework for relational data anomaly detection, integrating attribute and relational edge insights.
• Unsupervised Causal Abstractions Discovery — This work proposes an unsupervised method to discover high-level causal abstractions directly from low-level observational data.
• Universal Transformation of One-Class Classifiers for Unsupervised Anomaly Detection — A dataset folding method converts one-class classifiers into unsupervised anomaly detectors without modifying their core.
• Unlocking the relational power of graph neural networks in fraud prevention — Graph Neural Networks enhance fraud detection by modeling relational data, uncovering networked fraud invisible to traditional tabular methods.
• A Zero-shot Generalized Graph Anomaly Detection Framework via Node Reconstruction — AlignGAD enables zero-shot graph anomaly detection across diverse domains by unifying features and leveraging reconstruction discrepancy.
• The Data Points That Don’t Belong -And Why They Matter More Than You Think — Modern outlier detection leverages diverse techniques to find anomalies in complex, high-dimensional, and streaming datasets.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.LG updates on arXiv.org.