Year: 2026
Type: article
Abstract: Anti-money laundering (AML) transaction detection faced scarce labels, extreme class imbalance, relational dependencies, and future-information leakage. This study developed a validation-gated ensemble detection framework combining temporally constrained transaction features, historical account statistics, explicit transaction-network motifs, and seven heterogeneous XGBoost experts. A large language model (LLM) proposed candidate rules from aggregated validation evidence, which a programmatic rule-validation agent assessed through threshold fitting and out-of-time validation. A validation-gated router integrated expert predictions and validated rule signals. On the temporal test set of the synthetic HI-Small dataset, the framework achieved PR-AUC, F1, and recall of 0.5965, 0.5532, and 0.4299, respectively. These values exceeded those of the evaluated XGBoost and GIN-family baselines; the explicit-motif XGBoost baseline achieved a PR-AUC of 0.4788. Stage-wise analysis attributed most gains to explicit transaction-network motifs and validation-weighted expert voting. The router provided only a small improvement, and the rule branch did not improve PR-AUC. Under this experimental setting, the findings supported using the LLM for candidate-rule generation and auxiliary analysis, but did not establish it as the source of the overall performance gain.
Author: Yunzhe Wan
Institution: Dongbei University of Finance and Economics
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7235178359
