Robustness to Missing Samples in RF Classification Ensembles: NaN Sanitation Strategies Compared

Robustness to Missing Samples in RF Classification Ensembles: NaN Sanitation Strategies Compared

Explainable and Unbreakable RF Ensembles

Vote Tracing + Missing-Sample Robustness

Modern RF scenes are messy: bursts of dropouts, weird emitters you’ve never trained on, and models that sometimes disagree for reasons that only show up under pressure. We fixed all three.

• Explainability, for real: Every decision is auditable with exact Shapley (permutation) attributions on by default—no sampling noise, no knobs. You get per-model probabilities, contribution scores, vote timelines, pairwise disagreement, OSR signals, and a paper-trail in signal.metadata for every classification.
• Open-set that wins the bake-off: Our Energy + Disagreement OSR beats ODIN/Mahalanobis/MOS in RF while adding zero extra forwards, memory, or train-time fitting. Mahalanobis + EVT support is available for apples-to-apples baselines—fitter + ROC generator included.
• Robust to missing samples: When inputs go NaN, linear interpolation preserves accuracy at low–moderate corruption, nan_to_num is the latency champ, and mask stats expose burstiness so you can choose policy by regime. This isn’t theory; we wired the sanitation hooks into both temporal and spectral builders and measured error, p50/p95 latency, PSD KL, and mask statistics across corruption/SNR.

Why this matters

• Ops trust: Exact per-model attributions and vote timelines turn “because the net said so” into an audit log your compliance folks can live with.
• Field stability: Sensor dropout doesn’t crater accuracy or crash inference; the sanitation path is explicit, measurable, and logged.
• No deployment tax: You keep line-rate throughput. OSR and explainability ride the same trace you already collect.

Download the papers

• Vote Tracing (Rev3): [/mnt/data/Vote Tracing Model-Level Explainability for RF Signal Classification Ensembles bgilbert1984 Rev3.pdf](/mnt/data/Vote Tracing Model-Level Explainability for RF Signal Classification Ensembles bgilbert1984 Rev3.pdf)
• Missing-Sample Robustness: [/mnt/data/Robustness to Missing Samples in RF Classification Ensembles NaN Sanitation Strategies Compared bgilbert1984.pdf](/mnt/data/Robustness to Missing Samples in RF Classification Ensembles NaN Sanitation Strategies Compared bgilbert1984.pdf)

One-minute quickstart (your existing repos/targets)

# 1) Vote Tracing: figures + OSR baselines
cd /home/bgilbert/paper_Explainability_from_Vote_Traces
make xai-figs           # vote timeline (correct/incorrect), Shapley bars, disagreement heatmap
make osr-all            # fit Mahalanobis(+EVT) + generate ROCs
make pdf                # camera-ready PDF

# 2) Missing-Sample Robustness: full sweep + tables/figs
cd /home/bgilbert/paper_NaN_Padding_Interpolation_Robustness
make dev-quick          # runs corruption sweep, renders tables, builds PDF
# SNR-stratified variant:
python3 scripts/corruption_robustness.py --snr-bins "-10,-5,0,5,10,15" --pad-edges
python3 scripts/render_tables_mask_stats.py --global-json data/robustness_metrics.json \
  --snr-json data/robustness_metrics_snr.json --out tables/robustness_mask_tables.tex --focal_ratio 0.2
make pdf

What to show in the blog post (visuals)

1. Vote timeline (one correct 64QAM @ +15 dB, one wrong 8PSK→16QAM @ 0 dB).
2. Shapley bar triples for the three hardest cases (negative contributors pop in red).
3. Disagreement heatmap over 50k samples—instantly reveals diversity vs. clones.
4. Error vs. corruption and latency vs. corruption; annotate the “policy switch” (≤20% use interp_lin for accuracy, otherwise fall back to nan_to_num if you’re latency-bound).

Where this sells (and why now)

• Defense & spectrum regulators: auditability + OSR without extra compute is procurement catnip.
• Industrial wireless & telecom ops: drop-tolerant inference under maintenance windows or flaky links.
• Vendors: license the explainability/OSR SDK and the sanitation hooks as a compliance + reliability add-on to existing AMC stacks.

Repository: bgilbert1984/Robustness-to-Missing-Samples-in-RF-Classification-Ensembles-NaN-Sanitation-Strategies-Compared: We quantify the impact of input sanitation strategies—nan_to_num, zero-padding, and linear interpolation—on classification error and latency under controlled NaN corruption of IQ streams. We integrate sanitation hooks in temporal and spectral feature builders and systematically evaluate robustness across corruption ratios.