Optimize SMS spam-detection ensembles for realistic prevalence and asymmetric costs

Tune the ensemble configuration for SMS spam detection, particularly higher-recall larger minority-vote ensembles, subject to an explicit false-positive-rate constraint under realistic spam prevalence, so that deployment performance reflects the greater operational cost of false positives than false negatives.

Background

The paper’s class-imbalance analysis holds each classifier’s operating point fixed and analytically recomputes precision at several spam prevalences. Although this demonstrates degradation in positive predictive value as spam becomes rarer, it does not re-optimize the classifiers or ensembles for deployment conditions in which false positives are substantially more costly than false negatives.

The authors therefore identify prevalence-aware, cost-sensitive ensemble tuning as future work. The concrete unresolved task is to determine how ensemble size and voting rule should be selected under realistic spam prevalence and an explicit false-positive constraint, especially for larger minority-vote ensembles that achieve high recall at the cost of increased false positives.