Abstract
Pattern mining returns human-readable rules that analysts act on outside the sample that produced them, yet real data pool heterogeneous environments — sites, months, cohorts — so a rule that wins on pooled association may encode an environment-specific shortcut; every symbolic miner that penalizes a dispersion measures it within one sample, across resamples that are exchangeable by construction, and so cannot see shift. We propose SCPM, which estimates each candidate conjunction's risk-difference contrast separately inside every environment and ranks by mean contrast minus a cross-environment dispersion penalty — corrected, because the observed spread of those estimates is largely their own sampling error, so we penalize instead the random-effects heterogeneity that survives the noise floor. Branch-and-bound returns the exact top-K ranking under an admissible bound derived for that corrected score. We evaluate on eleven public datasets (66 leave-one-environment folds) spanning clinical, e-commerce, recommendation and geographic sources, comparing twelve ranking rules over one shared candidate pool under an explicit environment-leakage protocol. All seven environment-aware criteria rank ahead of all five within-sample criteria without overlap; SCPM lowers held-out-inclusive contrast dispersion against pooled ranking on all eleven datasets and 65 of 66 folds (p=0.001), and lowers transfer error at the unseen environment on 9 of 11 (p=0.019). Removing the noise floor leaves the score far less sensitive to its weight, retaining 78% of its mean contrast at a sixfold overshoot against 57% uncorrected, and its selections are insensitive to the heterogeneity estimator. SCPM makes cross-environment stability a practical screen wherever readable rules must travel.