PerCom Writing Style
Use this when revising the main paper. PerCom papers are IEEE Xplore proceedings read by ubicomp
reviewers, so they need a pervasive-computing contribution stated in the first page and
evidence a reviewer trusts on people they did not train on. The failure this skill prevents is
a technically fine paper that reads like an ML result with a wearable title glued on, or a systems
demo with no human at its center.
Revision rules
- Lead with the ubicomp contribution: the problem a real user or deployment faces, why current
sensing is inadequate, the contribution (system and/or finding), the evidence, and what changes
for pervasive computing.
- State claims at the right granularity. A recognition claim must say cross-subject or
within-subject, on what population, with which metric. "97% accuracy" without a split or a class
balance is a red flag to a PerCom reviewer, not a headline.
- Pair every claim with proportional evidence — real subjects, a fair baseline, F1 with
confidence intervals on realistic class balance, deployment realism — not adjectives.
- Argue limitations; do not recite them. Name the external, construct, and generalization
limits that actually bite this study (subject diversity, ground-truth quality, lab vs.
free-living), and say what you did to bound each. A boilerplate limitations paragraph tells a
reviewer you have not stressed your own claims.
- Respect the 9-page budget as a design constraint, not a formatting afterthought — IEEEtran
two-column is tight, and a study that only fits by shrinking the evaluation or limitations is
over-scoped. Recover space editorially, never by touching the template.
- Maintain double-blindness in self-citations (third person), testbed and system names, dataset
links, acknowledgements, and funding.
Ubicomp paper skeleton
| Section |
Job it must do |
Common failure |
| Intro |
Problem, inadequacy, contribution, evidence preview, ubicomp payoff — first page |
Leads with a technology trend, not a real-use problem |
| Background/Motivation |
Why a user or deployment needs this now |
Motivation by assertion, no grounding in practice |
| System / Study design |
The technique or the study + sensing protocol, reproducibly |
Method or sensor setup described too thinly to re-run |
| Evaluation |
Each claim answered with cross-subject, proportional evidence |
Within-subject or pooled-accuracy metrics that flatter the result |
| Limitations |
The limits that bite, each bounded |
Generic list untethered from this study's subjects/sensors |
| Related work |
Delta-first positioning against the ubicomp literature |
Catalog of citations with no contrast |
Sentence-level rewrites
| Draft pattern |
PerCom-safe rewrite |
| "Our system achieves 97% accuracy." |
"leave-one-subject-out F1 of 0.xx (95% CI ...) on participants" |
| "We evaluate on a large dataset." |
"We evaluate on participants over of free-living data, released as a dataset" |
| "Results show our approach works well." |
"cross-subject F1 improves by X over ; per-subject variance in Fig. 3" |
| "State-of-the-art performance." |
Claim scoped to the subjects, sensors, and setting actually tested |
| "The model recognizes activities." |
"the recognizer reaches F1 0.xx on held-out subjects for " |
Cross-subject and metric discipline
[Split] state within-subject vs. leave-one-subject-out (or session-out); PerCom default is cross-subject
[Balance] report class balance; on imbalanced activities use F1 (macro + per-class), not raw accuracy
[Event vs frame] say whether metrics are frame-level or event-level; they can differ sharply
[Realism] lab vs. free-living; scripted vs. spontaneous behavior -- name which you tested
-> for each: state the choice next to the number so a reviewer is not left guessing
Vignette: compressing an over-length study into 9 pages
A draft with three activity classes, nine figures, and a sprawling background: keep the
cross-subject headline result, the two figures that carry it, a per-class F1 table, and a
limitations subsection tied to subject diversity and ground truth; move per-subject breakdowns and
extra ablations to the dataset with explicit forward references; cut background to what the argument
needs. The test of a good cut: a reviewer should be able to answer "does it work on a new person,
and what threatens that?" from the body alone.
Output format
[Writing diagnosis] clear / under-motivated / over-claimed / within-subject-only / over-scoped
[First-page fix] <new framing leading with the pervasive-computing contribution>
[Claim audit] <claim -> split (LOSO?) -> metric (F1?) -> where answered -> proportional? yes/no>
[Limitations fix] <limit that bites -> bounding to add, placed by the result>
[Anonymity edits] <system names / self-citations / dataset links to rewrite>
1---2name: percom-writing-style3description: Use when revising an IEEE PerCom paper for a pervasive-computing contribution on the first page, cross-subject claims stated precisely, a limitations section that argues rather than recites, evidence proportional to the claim, double-blind wording, and disciplined use of the tight IEEEtran 9-page budget.4---56# PerCom Writing Style78Use this when revising the main paper. PerCom papers are IEEE Xplore proceedings read by ubicomp9reviewers, so they need a **pervasive-computing contribution stated in the first page** and10evidence a reviewer trusts on **people they did not train on**. The failure this skill prevents is11a technically fine paper that reads like an ML result with a wearable title glued on, or a systems12demo with no human at its center.1314## Revision rules1516- **Lead with the ubicomp contribution:** the problem a real user or deployment faces, why current17 sensing is inadequate, the contribution (system and/or finding), the evidence, and what changes18 for pervasive computing.19- **State claims at the right granularity.** A recognition claim must say **cross-subject or20 within-subject**, on what population, with which metric. "97% accuracy" without a split or a class21 balance is a red flag to a PerCom reviewer, not a headline.22- **Pair every claim with proportional evidence** — real subjects, a fair baseline, F1 with23 confidence intervals on realistic class balance, deployment realism — not adjectives.24- **Argue limitations; do not recite them.** Name the external, construct, and generalization25 limits that actually bite *this* study (subject diversity, ground-truth quality, lab vs.26 free-living), and say what you did to bound each. A boilerplate limitations paragraph tells a27 reviewer you have not stressed your own claims.28- **Respect the 9-page budget as a design constraint,** not a formatting afterthought — IEEEtran29 two-column is tight, and a study that only fits by shrinking the evaluation or limitations is30 over-scoped. Recover space editorially, never by touching the template.31- **Maintain double-blindness** in self-citations (third person), testbed and system names, dataset32 links, acknowledgements, and funding.3334## Ubicomp paper skeleton3536| Section | Job it must do | Common failure |37|---|---|---|38| Intro | Problem, inadequacy, contribution, evidence preview, ubicomp payoff — first page | Leads with a technology trend, not a real-use problem |39| Background/Motivation | Why a user or deployment needs this now | Motivation by assertion, no grounding in practice |40| System / Study design | The technique or the study + sensing protocol, reproducibly | Method or sensor setup described too thinly to re-run |41| Evaluation | Each claim answered with **cross-subject**, proportional evidence | Within-subject or pooled-accuracy metrics that flatter the result |42| Limitations | The limits that bite, each bounded | Generic list untethered from this study's subjects/sensors |43| Related work | Delta-first positioning against the ubicomp literature | Catalog of citations with no contrast |4445## Sentence-level rewrites4647| Draft pattern | PerCom-safe rewrite |48|---|---|49| "Our system achieves 97% accuracy." | "leave-one-subject-out F1 of 0.xx (95% CI ...) on <N> participants" |50| "We evaluate on a large dataset." | "We evaluate on <N> participants over <duration> of free-living data, released as a dataset" |51| "Results show our approach works well." | "cross-subject F1 improves by X over <baseline>; per-subject variance in Fig. 3" |52| "State-of-the-art performance." | Claim scoped to the subjects, sensors, and setting actually tested |53| "The model recognizes activities." | "the recognizer reaches F1 0.xx on held-out subjects for <activity set>" |5455## Cross-subject and metric discipline5657```text58[Split] state within-subject vs. leave-one-subject-out (or session-out); PerCom default is cross-subject59[Balance] report class balance; on imbalanced activities use F1 (macro + per-class), not raw accuracy60[Event vs frame] say whether metrics are frame-level or event-level; they can differ sharply61[Realism] lab vs. free-living; scripted vs. spontaneous behavior -- name which you tested62-> for each: state the choice next to the number so a reviewer is not left guessing63```6465## Vignette: compressing an over-length study into 9 pages6667A draft with three activity classes, nine figures, and a sprawling background: keep the68cross-subject headline result, the two figures that carry it, a per-class F1 table, and a69limitations subsection tied to subject diversity and ground truth; move per-subject breakdowns and70extra ablations to the dataset with explicit forward references; cut background to what the argument71needs. The test of a good cut: a reviewer should be able to answer "does it work on a new person,72and what threatens that?" from the body alone.7374## Output format7576```text77[Writing diagnosis] clear / under-motivated / over-claimed / within-subject-only / over-scoped78[First-page fix] <new framing leading with the pervasive-computing contribution>79[Claim audit] <claim -> split (LOSO?) -> metric (F1?) -> where answered -> proportional? yes/no>80[Limitations fix] <limit that bites -> bounding to add, placed by the result>81[Anonymity edits] <system names / self-citations / dataset links to rewrite>82```
brycewang-stanford