[ellispatrick]

Part 9 of the stack (#126 → #133 → this). Based on #133. It can be accepted or rejected on its own: closing it keeps glmnet's default lambda path.

Summary

Changes results (decided by Ellis Patrick's delegation, on the evidence below). The Cox elastic net (classifier = "CoxNet") ends its path of lambda at 0.05 of the largest value by default. glmnet's own default is 0.0001 with more samples than features, 0.01 otherwise. Tuning by C-index is kept.

Why

• Statistical: lambdas near zero give nearly unpenalised Cox models with dozens of coefficients. With the few events typical of omics cohorts (39 relapses in the protocol's METABRIC data), these overfit, and glmnet converges slowly there (it reported non-convergence at the end of the default path).
• Precedent: ncvreg ends its path at 0.05 when there are more features than samples.
• Performance: no loss on the protocol's own analysis; see below.
• Speed: 6.6× faster on that analysis.

Evidence

Protocol Procedure 2 step 7: merge of clinical, Domain Proportions, Proportions of Parent, Colocalisation and RNA Microarray views; all 31 combinations; 20 × 5 CV; shared folds; 16 cores.

Setting	Tuning	Time	Mean C-index over 31 combinations	Best combination
A (glmnet default path)	C-index	9.2 min	0.599	0.670
B	deviance	9.1 min	0.587	0.646
C (this PR)	C-index, path ends at 0.05	1.4 min	0.602	0.672
D	deviance, path ends at 0.05	1.4 min	0.587	0.647
E	deviance, path stops at 30 features	1.7 min	0.587	0.646

• Tuning measure: partial likelihood deviance is the textbook way to tune a Cox penalty, but it gave consistently lower C-indexes (−0.012), so C-index tuning stays.
• The cut at 0.05: with glmnet's default path, cross-validation chose a lambda below 0.05 of the largest in 12% of the 3,100 fold models (median 0.17). Clipping those did not lower the C-index; setting C was best in 21 of 31 combinations.
• Best combination: the same under every setting, clinical + Domain Proportions + Colocalisation.
• Caveat: the setting was chosen partly on this data set. The C-index difference is within noise, so the case rests on "no loss, six times faster, statistically sensible".

What changes

• coxnetTrainInterface passes lambda.min.ratio = 0.05 unless the user gives lambda or lambda.min.ratio (e.g. extraParams = list(train = list(lambda.min.ratio = 1e-4)) restores glmnet's default).
• A NEWS entry and a test.

Evidence

544 test expectations pass. Scripts and per-combination results: ProjectClassifyR scripts/coxnet_tuning/.

🤖 Generated with Claude Code

https://claude.ai/code/session_01TfsjwkbDbHkX4kJZNe4ER4

[DarioS]

Speed improvement but the same accuracy.