Geoffrey J. McLachlan

Research profile

Research summary

Geoffrey J. McLachlan is an Australia-based statistician at Queensland University of Technology whose OpenAlex-resolved profile shows a long, tightly focused record in mixture models and statistical learning. The snapshot reports an h-index of 67 and an i10-index of 212, with year-by-year counts running from 1972 through 2026 — more than five decades of nearly continuous publication. Topic mass sits in Bayesian methods and mixture models (321 works), statistical methods and applications (179), statistical distribution estimation (137), advanced clustering algorithms (133) and gene-expression cancer classification (77). Activity remains strong late in the record: 26 works in the most recent three years, produced with a compact circle of 4 distinct coauthors. Evolve has reviewed two 2026 arXiv preprints on which McLachlan is the last-listed author. One derives active-learning label acquisition from zero-one classification-risk curvature — judged rigorous and honest, with thin empirical gains over entropy and margin baselines. The other shows that informative MAR label missingness can beat fully classified learning in Weibull mixture models — judged theoretically sound, with convincing simulations but a semi-synthetic real-data test. A third 2026 preprint on information geometry and excess risk awaits review. The output and impact radar axes are not yet scored in this snapshot.

Research keywords

Highlights

• Risk-curvature theory for active label acquisition

  Evolve's review of this 2026 arXiv preprint (last author) finds the derivation of label choice from zero-one classification-risk curvature rigorous and honest, but judges the empirical gains over entropy and margin baselines thin; reproducibility is rated medium.

• Informative MAR labels can help in Weibull mixtures

  In this 2026 arXiv preprint (last author), informative missing-at-random labels are shown to beat fully classified learning in Weibull mixture models; the theory is judged sound, simulations convincing, but the real-data test is only semi-synthetic. Reproducibility is rated medium.

• Information geometry of missing labels in multiclass classification

  A third 2026 arXiv preprint with McLachlan as last author extends the missing-label line to information geometry and excess risk for multiclass classification; it has not yet been reviewed by Evolve, so no verdict, keywords or repro-readiness rating are available.

Timeline

1. 1972Earliest year in the countsByYear series, with 2 works recorded.
2. 1992A pronounced peak of 22 works, the highest annual count in the first two decades of the record.
3. 200025 works, an early-2000s high matched again in 2007.
4. 201635 works, the highest single-year count in the entire 1972–2026 series.
5. 2026Three arXiv preprints on active learning and missing-label classification enter the Evolve catalog, all with McLachlan as last author; the countsByYear series lists 19 works for this year.

Papers in the Evolve catalog

• 2026-09-06

  Large Classification-Risk-Optional Label Acquisition

• 2026-09-01

  Semi-Supervised Classification with Informative Missing Labels in Weibull Mixture Models

Frequent co-authors