Predicting Hard Drive Failures Using Mondrian Conformal Prediction

This study explores how Mondrian Conformal Prediction (MCP) enhances traditional k-Nearest Neighbors (kNN) models in predicting hard drive failures. Using Baidu’s open-source dataset of over 23,000 Seagate HDDs, the experiment demonstrates that MCP increases accuracy for detecting failing disks, despite dataset imbalance. More importantly, it enables the selective scrubbing of only 22.7% of drives — drastically cutting energy use while maintaining reliability. The results highlight the value of confidence scoring in large-scale predictive maintenance systems.