[ellispatrick]

Part of a stack of 8 pull requests from the October 2026 audit of ClassifyR (AUDIT.md on branch audit-2026-10). Each part is based on the one before, so they need to be merged in order. Each one can be accepted, changed or rejected on its own merits. Draft PR #125 holds all eight together. Every part installs and passes its own tests.

Stack, merge in order: #126 tests and core fixes → #127 interface fixes → #128 metrics and plots → #129 pathways and criss-cross → #130 maintenance → #131 faster cross-validation (identical results) → #132 shared folds (changes results) → #133 release notes. The regression harness and timings are in SydneyBioX/ProjectClassifyR (private).

Summary

Fixes to the classifier, feature ranking and helper interfaces: wrong results, and code paths that failed on any input.

Why and what (one commit each, each with a test that failed before)

Wrong results with no error:

• DLDA: the class prior had the wrong sign. It added log(prior) to a score that is minimised, which favoured the smaller class: on pure noise with 90/10 classes, 97% of samples were predicted to be the minority class. It now subtracts 2 log(prior), and posteriors are computed on the log scale (they underflowed with many features).
• GLM: fitted without an intercept (class ~ . + 0), so a shifted feature gave chance accuracy. Now fitted with an intercept.
• CoxNet feature ranking used the signed coefficient, so protective features ranked below features with zero coefficients. Now uses the absolute value.
• CoxNet, penalised GLM and XGB encoded test data separately from training data. A factor level missing from a test fold gave an error, or silently misaligned columns. Test data are now encoded with the training columns and levels.

Failed on any input:

• XGB with xgboost 3: ported to xgb.DMatrix/xgb.train.
• SVM with factor features.
• kNN returned characters, and failed for one test sample.
• Fisher discriminant, k-TSP and Poisson LDA referred to an undefined trainingMatrix.
• Naive Bayes and mixtures of normals with "crossover distance" weighting.
• colCoxTests with one feature, and its "slow" option.
• subtractFromLocation with one feature.
• Prevalidation with more than two classes.
• limmaRanking's ....
• edgesToHubNetworks with a matrix.

Surprising behaviour:

• Prevalidation and PCA seeded their inner CV with .Random.seed[1] (the RNG kind code, the same in every fold); they now draw a seed.
• previousSelection's overlap check could never fire, and now warns.
• KS, KL and pairs-differences rankings silently used only the first two classes, and now stop.

Speed, with identical results:

• Levene ranking uses one F-test over all features: the same order as car::leveneTest, 36.8 s → 2.1 s on asthma.
• Penalised GLM chooses lambda from one predict call: the same choice.

Do results change?

Yes, where a fix applies: DLDA predictions with unbalanced classes, GLM predictions, CoxNet rankings and XGB (which works now). Rankings and selections of the other methods are unchanged (regression harness).

Evidence

157 tests pass. Each commit's message gives its test.

Questions (left unchanged on purpose)

• xgboost 3 estimates base_score from the data.
• The penalised GLM lambda filter uses the signed sum of the first class's coefficients.
• previousSelection applies make.names, so non-syntactic names never match.
• The naive Bayes crossover uses unscaled densities.
• The likelihood-ratio ranking has a 25× faster closed form, but it would reorder two features whose statistic is exactly 0 (ties), so it is not included.

🤖 Generated with Claude Code

https://claude.ai/code/session_01TfsjwkbDbHkX4kJZNe4ER4

[DarioS]

Some good catches.