Location
https://www.kennesaw.edu/ccse/events/computing-showcase/fa25-cday-program.php
Document Type
Event
Start Date
24-11-2025 4:00 PM
Description
Fraud models routinely flag suspicious transactions but rarely explain why, which slows investigations and erodes trust. In this work we study Extended Isolation Forest (EIF) for unsupervised fraud detection and propose STAR-CAST, a lightweight framework that turns raw anomaly scores into threshold-aligned IF–THEN rule cards with explicit reliability measures. Using the public credit-card fraud dataset (284,807 transactions, 492 frauds; ~0.17% prevalence), we apply a time-aware 70/15/15 Train/Validation/Test split and fit-on-train preprocessing (Amount log1p→z; Time z; V1–V28 retained). We train IF, EIF, an EIF ensemble, a Mahalanobis baseline, and density models (HBOS, COPOD, ECOD) fully unsupervised, evaluate them as rankers first (PR-AUC, Max-F1, Precision@K), then freeze Validation thresholds and measure Test-set performance at FPR-targeted and recall-targeted operating points, with hourly block-bootstrap CIs to capture uncertainty. On Test, density/tail methods perform best on these PCA-style features (HBOS AP ≈ 0.139, COPOD ≈ 0.103, ECOD ≈ 0.072), while EIF consistently improves over IF (EIF AP ≈ 0.0534, EIF-ENS ≈ 0.0537, IF ≈ 0.0498; Mahalanobis ≈ 0.0456). At a realistic FPR of 0.5%, HBOS and COPOD achieve usable precision and recall, while EIF remains operationally competitive. STAR-CAST generates compact rules with high stability, fidelity to the model decision, and validated local precision, providing auditable, actionable explanations rather than opaque scores.
Included in
GC-0251 Reproducing Extended Isolation Forests with STAR-CAST
https://www.kennesaw.edu/ccse/events/computing-showcase/fa25-cday-program.php
Fraud models routinely flag suspicious transactions but rarely explain why, which slows investigations and erodes trust. In this work we study Extended Isolation Forest (EIF) for unsupervised fraud detection and propose STAR-CAST, a lightweight framework that turns raw anomaly scores into threshold-aligned IF–THEN rule cards with explicit reliability measures. Using the public credit-card fraud dataset (284,807 transactions, 492 frauds; ~0.17% prevalence), we apply a time-aware 70/15/15 Train/Validation/Test split and fit-on-train preprocessing (Amount log1p→z; Time z; V1–V28 retained). We train IF, EIF, an EIF ensemble, a Mahalanobis baseline, and density models (HBOS, COPOD, ECOD) fully unsupervised, evaluate them as rankers first (PR-AUC, Max-F1, Precision@K), then freeze Validation thresholds and measure Test-set performance at FPR-targeted and recall-targeted operating points, with hourly block-bootstrap CIs to capture uncertainty. On Test, density/tail methods perform best on these PCA-style features (HBOS AP ≈ 0.139, COPOD ≈ 0.103, ECOD ≈ 0.072), while EIF consistently improves over IF (EIF AP ≈ 0.0534, EIF-ENS ≈ 0.0537, IF ≈ 0.0498; Mahalanobis ≈ 0.0456). At a realistic FPR of 0.5%, HBOS and COPOD achieve usable precision and recall, while EIF remains operationally competitive. STAR-CAST generates compact rules with high stability, fidelity to the model decision, and validated local precision, providing auditable, actionable explanations rather than opaque scores.