Graph-Based Anomaly Detection for Data Integrity and Fraud Prevention in Warehouse Management Systems: A PRISMA-Guided Systematic Literature Review
Keywords:
Graph Anomaly Detection, Warehouse Data Integrity, Fraud Prevention, Inventory Record Inaccuracy, Systematic ReviewAbstract
This study examined graph-based anomaly detection for data integrity and fraud prevention in warehouse management systems through a PRISMA-guided quantitative systematic literature review framework. It organized the literature by data source, graph representation, detection target, method family, learning paradigm, validation design, and reported performance measures. The proposed review process included searching academic databases, removing duplicate records, screening titles and abstracts, assessing full texts, and extracting study-level data. Because verified search and extraction results were unavailable, all numerical values in the current draft are examples rather than findings from a completed review. In that example, 1,268 records were identified, 318 duplicates were removed, and 30 studies were retained from 140 full-text reports assessed. The study profile included 8 studies using operational warehouse or inventory data, 12 using public graph benchmarks from adjacent domains, 2 using public warehouse or logistics data, 5 using simulated data, and 3 combining operational and simulated data. Heterogeneous static graphs represented 9 of the example studies, and node- or entity-level detection was the most frequent target, appearing in 12. Graph convolutional or attention networks formed the largest single method group, with 8 studies, while supervised learning was the most common learning paradigm, with 13. Precision and recall were each reported in 21 example studies, whereas confidence intervals were reported in 7. Random data splitting was the most frequent validation strategy, used in 12 studies; facility-based validation appeared in 2. These patterns show how the review can distinguish direct warehouse evidence from methods transferred from financial fraud, e-commerce, cybersecurity, and other fields. They also demonstrate the importance of separating inventory discrepancies and process deviations from suspected or confirmed fraud. Differences in data sources, labels, graph construction, validation, and performance metrics limit direct comparison across studies. The example values must be replaced with verified screening and extraction results before the abstract can represent completed research.


