[ellispatrick]

Part 17 (#140 → #141 → #142 → this). Based on #142 (the registry). The four classifiers are added through the registry. Each is in Suggests, and existing results are unchanged.

Summary

Four new classifier keywords, each one interface file and one registry entry:

Keyword	Library	Outcomes	Model
"aorsf"	aorsf	classes, survival	oblique random forest (linear-combination splits), 500 trees
"glmboost"	mboost	survival	componentwise boosted Cox model, mstop = 100, nu = 0.1
"LiblineaR"	LiblineaR	classes	logistic regression with an L2 (default) or L1 penalty; cost chosen by cross-validation
"ncvreg"	ncvreg	two classes, survival	MCP-penalised logistic or Cox regression (SCAD and lasso available); lambda chosen by cross-validation

Why

Ellis Patrick's decision on Q27–Q30 (add all four), ProjectClassifyR docs/DECISIONS.md, 10 October 2026. The survey in docs/CLASSIFIER_LIBRARIES.md found:

• aorsf gave the best forest C-index on METABRIC and fitted 14–80× faster than randomForestSRC.
• glmboost gave the best boosting C-index on Colocalisation.
• LiblineaR was the fastest and most accurate linear model on asthma.
• ncvreg offers MCP and SCAD penalties.

What changes

• R/interfaceObliqueForest.R:
  • One interface for both outcomes.
  • Class probabilities, or predicted mortality as the risk score.
  • Features are ranked by ANOVA importance.
  • It uses one thread by default (aorsf's own default is every core), and set.seed makes it reproducible.
• R/interfaceBoostedCox.R:
  • The linear predictor is the risk score, and the chosen coefficients give the selected features.
  • glmboost's warning that a Cox model has no intercept is muffled.
• R/interfaceLiblineaR.R:
  • Features are standardised by the training samples, and classes get inverse class size weights.
  • Without a cost, the cost is chosen from 10^(-3:2) by the balanced error of five-fold stratified cross-validation of the training samples (the smallest cost among ties).
  • Only the logistic types (0, 6, 7) are accepted, because they give class probabilities.
• R/interfaceNcvreg.R:
  • cv.ncvreg (binomial) or cv.ncvsurv, five folds, at lambda.min (ncvreg's default measure, the deviance).
  • More than two classes give a clear error.
• Registry entries: each has tuning presets for tuneParams = "auto", as listed in the presets vignette.
• Tests: the registry test fits and predicts each for every outcome it accepts. Further tests check reproducibility, that higher risk scores mean earlier events, the feature extractors, and the errors.

Evidence

20 × 5 cross-validation, shared folds, seed 1, 4 cores (ProjectClassifyR scripts/structure/new_classifiers_bench.R, results/structure/new_classifiers/tables.txt).

Asthma (190 samples, 124 vs 66), classes.

Features	Classifier	Time (s)	Balanced accuracy	AUC	Median features used
t-test top 20	DLDA	6.4	0.787	0.876	20
	randomForest	8.1	0.781	0.889	20
	aorsf	9.4	0.799	0.898	20
	elasticNetGLM	7.4	0.816	0.892	11.5
	LASSOGLM	7.7	0.810	0.892	8
	LiblineaR	5.9	0.810	0.894	20
	ncvreg	11.9	0.785	0.888	3
	SVM	5.3	0.776	0.874	20
	XGB	5.4	0.781	0.877	20
all 2000	randomForest	14.0	0.833	0.934	1643
	aorsf	51.4	0.829	0.937	630
	elasticNetGLM	14.9	0.909	0.976	85
	LiblineaR	24.7	0.942	0.989	2000
	ncvreg	21.9	0.864	0.947	8

METABRIC (165 patients, 39 relapses), relapse-free survival, CoxPH top 20 features of each view; C-index.

Classifier	clinical	RNA Microarray	Type Proportions	Type Protein Means	Colocalisation	Domain Proportions	Proportions of Parent	Time (s), all views
CoxPH	0.653	0.481	0.432	0.560	0.604	0.588	0.609	12.1
CoxNet	0.619	0.520	0.521	0.608	0.637	0.560	0.643	37.7
randomSurvivalForest	0.680	0.585	0.595	0.664	0.617	0.622	0.592	56.8
XGB	0.550	0.544	0.597	0.672	0.621	0.592	0.574	20.2
aorsf	0.661	0.564	0.565	0.655	0.642	0.603	0.622	30.9
glmboost	0.623	0.516	0.523	0.627	0.646	0.588	0.621	21.8
ncvreg	0.524	0.494	0.506	0.547	0.604	0.507	0.538	37.1

• aorsf: close to the best forest on every view and about half randomSurvivalForest's time. With all 2000 asthma features it is 3.7× slower than ranger, at the same accuracy.
• glmboost: the best C-index on Colocalisation, as in the survey.
• LiblineaR: the best asthma model with all features. With 20 features it matches the penalised GLMs and is the fastest of them.
• ncvreg: weak on METABRIC survival. With 39 events, its deviance-tuned MCP lambda gives very sparse models: 0–3 of the 20 Colocalisation features in 20 resampled training sets, sometimes none. It is offered as an option and isn't recommended as a default for small survival cohorts.
• Existing results: the regression harness (18 scenarios) is identical to 16/ One registry of classifiers and feature selection methods (no change to results) #142's build.
• Tests and check: R CMD check (tests, examples, vignettes) gives the same single NOTE as 15/ Defaults and dependencies: no rebalancing and serial by default; five packages to Imports, three dropped #141. Its test run passes 567 expectations, none failing.

🤖 Generated with Claude Code

https://claude.ai/code/session_01TfsjwkbDbHkX4kJZNe4ER4