Cost-Sensitive Learning for Rare Storage Failures: Evaluating Warning Lead Time, Missed Failures, and False-Alarm Burden

Abstract

Storage failures are rare relative to normal operating periods, so conventional classifiers may achieve high aggregate accuracy while missing operationally important events or generating excessive false alarms. We compared standard cross-entropy training with three cost-sensitive learning strategies on telemetry from 12,406 devices containing 418 documented failure events. Models were evaluated using time-ordered validation, with warning lead time, failure recall, precision, and alert burden measured at matched decision thresholds. Cost-sensitive training increased failure recall from 61% to 78% while preserving a median warning lead time of 9.6 days. The best-performing configuration reduced false alarms by 24% relative to class-weighting alone and generated 2.1 alerts per 100 device-days. Extremely aggressive miss penalties improved recall further but produced an alert volume unlikely to be operationally sustainable. Calibration deteriorated when cost weights were increased without post-training adjustment; isotonic recalibration corrected most of this effect. Cost-sensitive objectives can improve rare-failure detection, but model selection should explicitly balance missed failures, usable warning time, and the number of alerts an operations team can realistically investigate.