Supplementary material captions — Knowledge and Information Systems submission
Manuscript: Cost-Sensitive Incremental Thermal Prediction with Adaptive Drift Detection for Predictive Maintenance on Single-Board Computers
Authors: Abd. Hallim (corresponding, abd.halim@binus.ac.id), Maria Susan Anggreainy, Endra Oey, Widodo Budiharto
Affiliation: Bina Nusantara University, Jakarta, Indonesia

ESM_1.csv — Online Resource 1. Per-variant detection metrics and confusion matrices (TP/FP/FN/TN, recall, precision, F1, false-alarm rate, G-mean) for all fourteen CSL sweep variants (c ∈ {1,2,5,10,15,20,34} × ADWIN on/off). Source of Tables 1–2.

ESM_2.csv — Online Resource 2. Hoeffding Adaptive Tree (HAT) baseline results at c=1 and c=2, including footprint and latency. Source of the HAT rows in Table 3.

ESM_3.csv — Online Resource 3. Threshold-shifting baseline: detection metrics of HT-base with the alert boundary lowered across 56–60°C. Discussed in Section 5.4.

ESM_4.csv — Online Resource 4. Per-fold leave-one-device-out results across the eight LwHBench Raspberry Pi 4B units (HT-base vs HT-CS). Source of Table 5.