percom-writing-style

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 <N> participants" |

| "We evaluate on a large dataset." | "We evaluate on <N> participants over <duration> of free-living data, released as a dataset" |

| "Results show our approach works well." | "cross-subject F1 improves by X over <baseline>; 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 <activity set>" |

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>