Recalibrate RFF-MMD without the position-25 outlier

Determine how recalibrating the RFF-MMD threshold after excluding the position-25 natural-heterogeneity outlier affects detection performance on the Trendyol production benchmark.

Background

The RFF-MMD threshold is calibrated using the 99th percentile of pre-drift positions in the sudden–strong scenario. One pre-drift position, position 25, is an approximately 20-fold statistic outlier caused by natural source-table heterogeneity rather than injected drift. The paper retains this outlier in the reported calibration, leaving the effect of excluding it unresolved.

References

The quantile is dominated by the position-25 outlier (\approx 20\times the pre-drift mean; Section~(vi)), which we retain; recalibrating without it is left to future work.

— Beyond Marginal Monitoring: Distributed Joint-Distribution Testing for Data Concept Drift in Large Scale E-Commerce Operations  (2610.08132 - Pullu et al., 6 Oct 2026) in Section 6.3.1, Setup and threshold calibration

The same implementation reproduces published detection accuracies on Failing Loudly (Section~(ii)), suggesting the failure is scale-and-sampling-specific; a full mechanistic analysis at Trendyol scale is deferred to future work.

— Beyond Marginal Monitoring: Distributed Joint-Distribution Testing for Data Concept Drift in Large Scale E-Commerce Operations  (2610.08132 - Pullu et al., 6 Oct 2026) in Section 6.3.7, Failure modes: per-dim KS, MMD-RBF, and projected KS