Tabular In-Context Learning for Low-Resource Rare Industrial Fault Detection: a Cost-Sensitive Comparison on Scania APS and SECOM
Rare industrial fault detection is complicated by class imbalance, missing sensor values, asymmetric errors, and limited training resources. We compared TabICLv2 with TabM, XGBoost, and logistic regression on Scania APS Failure and SECOM. For Scania, nested stratified samples of 5,000 and 10,000 training records were drawn under five prespecified seeds. Shared three-fold splits supported out-of-fold (OOF) evaluation; tunable models were selected by OOF AUPRC, and decision thresholds were locked by minimizing OOF cost, defined as C = 10 × FP + 500 × FN. The official test set remained locked until all selections were complete. SECOM used three repeats of five-fold outer cross-validation with three-fold inner selection. Paired class-stratified bootstrap analyses used 2,000 replicates. At the prespecified 10,000-record Scania budget, TabICLv2 achieved an AUPRC of 0.8879, an official cost of 13,002, an MCC of 0.6514, and a Brier score of 0.00685; XGBoost achieved 0.8549, 16,456, 0.6125, and 0.00840, respectively. The TabICLv2-minus-XGBoost difference was 0.03294 for AUPRC (95% CI 0.02089–0.04570) and −3,454 for cost (95% CI −5,370.60 to −1,711.95). Across 15 SECOM outer folds, mean AUPRC was 0.2083 for TabICLv2 and 0.1705 for XGBoost. After averaging repeated OOF probabilities per record, the paired AUPRC difference was 0.05110 (95% CI −0.00693 to 0.11358). TabICLv2 required 992.2 s, approximately 413 times the XGBoost time. TabICLv2 improved the prespecified Scania outcomes but incurred substantial task-side computational cost; SECOM provided suggestive rather than definitive support.
