[ellispatrick]

Based on #134; independent of the other post-#134 PRs. It can be accepted or rejected on its own.

Summary

Changes results only for naive Bayes with weighting = "crossover distance" (4.7% of predictions on the asthma data). Every other change either fixes something that errored or leaves predictions unchanged: Fisher discriminant scores move but its classes don't, and weighted kNN changes only when a test sample is identical to a training sample. All 18 regression scenarios are identical to #134.

Why

Ellis Patrick decided these on 2026-10-10 (ProjectClassifyR DECISIONS.md, questions Q13, Q22–Q24, Q26):

• Q22, naive Bayes: vertical distances used class densities scaled by class size, but crossover points used the unscaled densities, so the two disagreed about where the classes meet whenever the classes differ in size.
• Q23, Fisher discriminant: the score was -aᵀx, not centred on the critical value. Its sign didn't match the predicted class, and scores from models trained on different folds sat on different offsets.
• Q23, kNN: a neighbour at distance zero has infinite weight, and that was replaced by an arbitrary 999999.
• Q24, previousSelection: it matched previous features after make.names(). Different names with the same safe name ("g-1" and "g.1") matched the wrong column; such data actually stopped with "duplicate row.names". It also needed the user to know to set intermediate = ".iteration".
• Q26, ensemble selection: the list-of-rankings branch of .doSelection used undefined variables (featuresLists, measurementsSubset, aResult), so it always failed.
• Q13, resubstitution tuning: when several nFeatures values fit the training set perfectly, the first is chosen silently. This is documented; the warning makes it visible.

What changes

• naiveBayesKernel: densities are scaled by class size once, and both the vertical distances and the crossover points use them. With weighting = "height difference" (the default) predictions are unchanged.
• fisherDiscriminant: score = critical value − aᵀx, positive for the second class. Predicted classes are unchanged.
• kNNinterface(mode = "weighted"): if any of the k neighbours is at distance zero, those neighbours alone vote, equally. Otherwise the weights are 1/distance as before.
• previousSelection matches on the original feature names (mcols(measurements)$feature, kept by prepareData), or on assay + feature for multi-assay data. SelectParams("previousSelection", ...) adds the .iteration intermediate itself.
• prepareData, runTest and predict make safe feature names unique (make.names(unique = TRUE)), so colliding names no longer stop the run. Names that were already unique are unchanged.
• Ensemble selection works: each ranking function ranks the features, and the selection is the features in the top n of at least minPresence rankings. With several nFeatures values, n is tuned by resubstitution or nested CV like the single-ranking path, and the tuning table is stored. otherParams can be one list per ranking function or shared by all of them.
• runTests warns once per cross-validation, giving the number of training sets in which resubstitution tuning tied at a perfect fit, and suggests tuneMode = "Nested CV".

Evidence

• Tests: 560 expectations pass, with new tests for each change (naive Bayes crossover against the analytic crossover, kNN exact match, Fisher score sign, colliding names in previousSelection, ensemble selection with and without tuning, the warning).
• Regression harness: all 18 scenarios identical to 9/9 Cox elastic net: lambda path ends at 0.05 of the largest value by default (changes results; 6.6x faster) #134.
• Changed classifiers, asthma data (500 features, t-test top 20, 5 × 5 CV, same seed):

Classifier	Predictions changed	Balanced accuracy	AUC
naive Bayes, crossover distance, weighted	4.7%	0.746 → 0.725	0.822 → 0.822
kNN, weighted, k = 5	0% (no exact duplicates)	0.761 → 0.761	0.868 → 0.868
Fisher discriminant	0%	0.777 → 0.777	pooled-score AUC 0.860 → 0.867

The naive Bayes change lowers balanced accuracy here. Scaling by class size moves the crossover towards the smaller class, so more samples near the boundary vote for the larger one. That is the consistent behaviour, but it is not an improvement on this data. Scripts: ProjectClassifyR scripts/classifier_fixes/.

R CMD check was not rerun for this PR; no documentation files changed.

🤖 Generated with Claude Code

https://claude.ai/code/session_01TfsjwkbDbHkX4kJZNe4ER4

[DarioS]

A couple of rare edge cases handled.