Per-person sequence models of keystroke behaviour: what remains of the gain against stronger comparisons

Whether a per-person model looks worthwhile depends on the comparison. Exploring public co-writing (CoAuthor) and programming (CS1) keystroke data, coded per second over overlapping 30-second windows, we scored models on each writer's later symbols. In CoAuthor, a per-writer three-state hidden Markov model beats a shared model of the same structure by up to 0.19 nats per symbol but loses to the writer's own first-order Markov chain. With a stronger structure and shrinkage chosen on other writers, about 5×10−4 remains at five and fifteen training sessions in the 15-session panel at the fixed base smoothing, no more than the shared model improves when its smoothing is removed; synthetic writers without per-person differences left a similar remainder relative to that improvement (ratio not prespecified). When the shared model's chosen smoothing is applied to both models with the structure fixed, 3.4×10−3 and 5.2×10−3 remain; this bears on a designer's decision but was not separated from closeness in time or the amount of data, nor checked synthetically. Neither value shows stable per-person differences; under the post hoc null at the fixed base smoothing, planted differences up to 3×10−2 were not reliably detected. The synthetic writers, both-sides smoothing and wider grids came after seeing results. In CS1, mostly pauses, the shared model is chosen at the fixed base smoothing. For overlapping-window symbols, group-level decisions on per-person modelling should compare with both baselines, choose structure and shrinkage on other people, and apply the shared model's chosen smoothing to both models (tried for one structure). This preprint is the manuscript before peer review, with the supplementary information (Online Resource 1, sections S1–S13) appended after the main text. The code, the conditions of every analysis and the aggregate and synthetic results are archived separately (see related identifiers).

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-21
DOI
https://doi.org/10.5281/zenodo.24402360
Primary Topic
User Authentication and Security Systems
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Per-person sequence models of keystroke behaviour: what remains of the gain against stronger comparisons

Mikto Miura
Zenodo (CERN European Organization for Nuclear Research)
User Authentication and Security Systems
preprint

Per-person sequence models of keystroke behaviour: what remains of the gain against stronger comparisons

Mikto Miura
preprint en

Abstract

Whether a per-person model looks worthwhile depends on the comparison. Exploring public co-writing (CoAuthor) and programming (CS1) keystroke data, coded per second over overlapping 30-second windows, we scored models on each writer's later symbols. In CoAuthor, a per-writer three-state hidden Markov model beats a shared model of the same structure by up to 0.19 nats per symbol but loses to the writer's own first-order Markov chain. With a stronger structure and shrinkage chosen on other writers, about 5×10−4 remains at five and fifteen training sessions in the 15-session panel at the fixed base smoothing, no more than the shared model improves when its smoothing is removed; synthetic writers without per-person differences left a similar remainder relative to that improvement (ratio not prespecified). When the shared model's chosen smoothing is applied to both models with the structure fixed, 3.4×10−3 and 5.2×10−3 remain; this bears on a designer's decision but was not separated from closeness in time or the amount of data, nor checked synthetically. Neither value shows stable per-person differences; under the post hoc null at the fixed base smoothing, planted differences up to 3×10−2 were not reliably detected. The synthetic writers, both-sides smoothing and wider grids came after seeing results. In CS1, mostly pauses, the shared model is chosen at the fixed base smoothing. For overlapping-window symbols, group-level decisions on per-person modelling should compare with both baselines, choose structure and shrinkage on other people, and apply the shared model's chosen smoothing to both models (tried for one structure). This preprint is the manuscript before peer review, with the supplementary information (Online Resource 1, sections S1–S13) appended after the main text. The code, the conditions of every analysis and the aggregate and synthetic results are archived separately (see related identifiers).

Zenodo (CERN European Organization for Nuclear Research)
User Authentication and Security Systems
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon