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>
1---2name: percom-experiments3description: Use when designing or auditing IEEE PerCom empirical evaluations, covering real human subjects, leave-one-subject-out / cross-subject evaluation, F1 and event-level metrics on imbalanced activity classes, deployment realism (free-living vs. lab), fair baselines, contamination-aware model ablations, and matching evidence to the shape of each pervasive-computing claim.4metadata:5 source: https://skillmd.com/skills/brycewang-stanford/percom-experiments6---78# PerCom Experiments910Use this before submission when the evaluation is not yet locked. PerCom reviewers are ubicomp11empiricists; the evaluation is where a sensing idea is won or lost, and — because the review is a12single round with a bounded rebuttal — **the evaluation must be complete at submission** (you13cannot add experiments in the rebuttal). The organizing principle is **evidence proportional to the14claim**, tested on **people and conditions a skeptic would accept**.1516## Evaluation audit1718- **Evaluate cross-subject by default.** A recognition claim about *users* needs19 **leave-one-subject-out** (or leave-one-session-out) results, not a pooled split that lets the20 same person appear in train and test. Within-subject numbers are a supporting detail, never the21 headline.22- **Use real human subjects,** described by count and relevant characteristics, with the collection23 protocol stated. Report how many, doing what, wearing/placed where.24- **Report the right metric.** Human activity is imbalanced (most of a day is "null"), so **F125 (macro and per-class)**, precision/recall, or **event-level** metrics tell the truth where raw26 accuracy flatters. Say whether metrics are frame-level or event-level.27- **Test deployment realism.** Distinguish **lab vs. free-living** and scripted vs. spontaneous28 behavior; a result only shown on scripted in-lab data invites the "does this survive daily life?"29 objection.30- **Choose fair baselines,** including the strongest prior method and a simple-but-reasonable31 alternative, tuned with a documented, equal budget. An untuned baseline is a scored weakness.32- **Report variance:** confidence intervals across subjects/folds, number of runs, and the source33 of stochasticity. Per-subject distributions matter more than a single mean here.34- **Design limitations in, not on:** know before you collect which generalization and construct35 limits the study will have (subject diversity, ground-truth quality), and instrument to bound36 them.3738## Claim-to-evidence design table3940| Ubicomp claim | Matching evidence | Reject pattern avoided |41|---|---|---|42| "Recognizes activity for new users" | Leave-one-subject-out F1 with per-subject spread | "Within-subject / pooled split inflates the number" |43| "Works in daily life" | Free-living data, event-level metrics | "Only scripted in-lab sessions tested" |44| "Beats the prior recognizer" | Same data + tuned baseline, equal budget | "Baseline untuned or on a different split" |45| "Handles class imbalance" | Macro-F1 + per-class recall, stated balance | "Raw accuracy hides the rare-class collapse" |46| "The model adds the value" | Ablation vs. classical features/heuristics | "Model's marginal contribution never isolated" |47| "Generalizes across contexts" | Diverse subjects/environments + explicit limits | "One population, claimed universal" |4849## Contamination- and leakage-aware evaluation5051Sensing pipelines leak in subtle ways; the reviewer's first questions are about splits and leakage:5253```text54[Subject leakage] never let one participant appear in both train and test -- LOSO prevents it55[Session/time leak] windows from one recording session can leak across a naive random split56[Normalization leak] fit scalers/PCA on train only; a global normalization leaks test statistics57[Pretraining] if a foundation model is used, report whether test subjects/data could be in its58 training set; prefer held-out or post-cutoff data59[Ablation] isolate the model's marginal value against a classical-feature baseline60```6162## Human-subjects provenance floor6364- State subject count and relevant demographics, the collection protocol, and IRB/consent status.65- Pin device models, firmware, sampling rates, and sensor placement; archive the **extracted,66 de-identified** dataset, not just a description.67- Report the labeling protocol, who labeled, and inter-annotator agreement — silent label noise68 skews every downstream number.6970## Vignette: evaluating a HAR recognizer7172Suppose the paper claims a wearable recognizer beats a prior model on daily activities. The matching73plan: collect from a diverse participant set over multiple days of free-living; evaluate74**leave-one-subject-out**; report macro-F1 and per-class recall with confidence intervals across75subjects; run both models on the same folds with an equal, documented tuning budget; add an ablation76against classical features; and state external validity (population, device) as a bounded limitation77— every number traceable to a logged run in the artifact, because the rebuttal cannot add a run.7879## Reporting floor8081- Cross-subject metric (F1/event-level) with confidence intervals and per-subject spread for the82 headline comparison; say what the intervals represent.83- Number of runs and the source of variance for any stochastic component.84- The compute actually consumed, not vague feasibility language.8586## Output format8788```text89[Evaluation readiness] strong / adequate / weak (remember: no new experiments in the rebuttal)90[Claim -> evidence map] <claim: subjects / split (LOSO?) / metric (F1?) / setting (free-living?)>91[Baseline fairness] <baseline -> tuned? equal budget? same split? documented?>92[Leakage check] <subject / session / normalization / pretraining leakage handled? yes/no>93[Limitations-by-design] <generalization/construct limit -> instrumentation to bound it>94[Decision-critical run to finish before submission] <one experiment>95```