percom-experiments

PerCom Experiments

Use this before submission when the evaluation is not yet locked. PerCom reviewers are ubicomp

empiricists; the evaluation is where a sensing idea is won or lost, and — because the review is a

single round with a bounded rebuttal — the evaluation must be complete at submission (you

cannot add experiments in the rebuttal). The organizing principle is **evidence proportional to the

claim, tested on people and conditions a skeptic would accept**.

Evaluation audit

• Evaluate cross-subject by default. A recognition claim about users needs

leave-one-subject-out (or leave-one-session-out) results, not a pooled split that lets the

same person appear in train and test. Within-subject numbers are a supporting detail, never the

headline.

• Use real human subjects, described by count and relevant characteristics, with the collection

protocol stated. Report how many, doing what, wearing/placed where.

• Report the right metric. Human activity is imbalanced (most of a day is "null"), so **F1

(macro and per-class), precision/recall, or event-level** metrics tell the truth where raw

accuracy flatters. Say whether metrics are frame-level or event-level.

• Test deployment realism. Distinguish lab vs. free-living and scripted vs. spontaneous

behavior; a result only shown on scripted in-lab data invites the "does this survive daily life?"

objection.

• Choose fair baselines, including the strongest prior method and a simple-but-reasonable

alternative, tuned with a documented, equal budget. An untuned baseline is a scored weakness.

• Report variance: confidence intervals across subjects/folds, number of runs, and the source

of stochasticity. Per-subject distributions matter more than a single mean here.

• Design limitations in, not on: know before you collect which generalization and construct

limits the study will have (subject diversity, ground-truth quality), and instrument to bound

them.

Claim-to-evidence design table

| Ubicomp claim | Matching evidence | Reject pattern avoided |

|---|---|---|

| "Recognizes activity for new users" | Leave-one-subject-out F1 with per-subject spread | "Within-subject / pooled split inflates the number" |

| "Works in daily life" | Free-living data, event-level metrics | "Only scripted in-lab sessions tested" |

| "Beats the prior recognizer" | Same data + tuned baseline, equal budget | "Baseline untuned or on a different split" |

| "Handles class imbalance" | Macro-F1 + per-class recall, stated balance | "Raw accuracy hides the rare-class collapse" |

| "The model adds the value" | Ablation vs. classical features/heuristics | "Model's marginal contribution never isolated" |

| "Generalizes across contexts" | Diverse subjects/environments + explicit limits | "One population, claimed universal" |

Contamination- and leakage-aware evaluation

Sensing pipelines leak in subtle ways; the reviewer's first questions are about splits and leakage:

[Subject leakage]  never let one participant appear in both train and test -- LOSO prevents it
[Session/time leak] windows from one recording session can leak across a naive random split
[Normalization leak] fit scalers/PCA on train only; a global normalization leaks test statistics
[Pretraining]      if a foundation model is used, report whether test subjects/data could be in its
                   training set; prefer held-out or post-cutoff data
[Ablation]         isolate the model's marginal value against a classical-feature baseline

Human-subjects provenance floor

• State subject count and relevant demographics, the collection protocol, and IRB/consent status.
• Pin device models, firmware, sampling rates, and sensor placement; archive the **extracted,

de-identified** dataset, not just a description.

• Report the labeling protocol, who labeled, and inter-annotator agreement — silent label noise

skews every downstream number.

Vignette: evaluating a HAR recognizer

Suppose the paper claims a wearable recognizer beats a prior model on daily activities. The matching

plan: collect from a diverse participant set over multiple days of free-living; evaluate

leave-one-subject-out; report macro-F1 and per-class recall with confidence intervals across

subjects; run both models on the same folds with an equal, documented tuning budget; add an ablation

against classical features; and state external validity (population, device) as a bounded limitation

— every number traceable to a logged run in the artifact, because the rebuttal cannot add a run.

Reporting floor

• Cross-subject metric (F1/event-level) with confidence intervals and per-subject spread for the

headline comparison; say what the intervals represent.

• Number of runs and the source of variance for any stochastic component.
• The compute actually consumed, not vague feasibility language.

Output format

[Evaluation readiness] strong / adequate / weak (remember: no new experiments in the rebuttal)
[Claim -> evidence map] <claim: subjects / split (LOSO?) / metric (F1?) / setting (free-living?)>
[Baseline fairness] <baseline -> tuned? equal budget? same split? documented?>
[Leakage check] <subject / session / normalization / pretraining leakage handled? yes/no>
[Limitations-by-design] <generalization/construct limit -> instrumentation to bound it>
[Decision-critical run to finish before submission] <one experiment>