[ellispatrick]

Part 16 (#140 → #141 → this → #143). Based on #141.

Summary

No change to results. Every classifier and feature selection method now has one entry in a registry, R/registry.R. Adding a method takes:

• one interface file;
• its Collate line;
• one registry entry.

It used to touch up to nine places.

Why

Ellis Patrick's decision on Q36 ("Yes"), ProjectClassifyR docs/DECISIONS.md, 10 October 2026, following docs/MAINTAINABILITY.md section 6. Before this, adding a classifier meant editing:

• the interface file and Collate;
• in constants.R: functionsTable and classifyKeywords;
• in utilities.R: .classifierKeywordToParams;
• simpleParams.R, for a *params() function and its tuning presets;
• cleanClassifier's hard-coded penalised-GLM list;
• two vignette tables.

Nothing recorded whether a method accepts classes, survival or both.

What changes

• R/registry.R: .registerClassifier() and .registerSelection() entries, built in .onLoad, so file order doesn't matter. Each entry gives:
  • the functions;
  • the display name;
  • the description;
  • the outcomes it accepts;
  • tuning presets and default parameters;
  • the fallback classifier for too few features (the penalised GLMs fall back to GLM).
• Read from the registry:
  • available()'s tables;
  • the keyword-to-function and keyword-to-parameters mappings;
  • the names shown in results and plots;
  • the too-few-features fallback.
• Removed:
  • simpleParams.R;
  • the keyword and display-name tables in constants.R.
• available() gains an Outcomes column (classes, survival or both).
• crossValidate stops at once with a clear error when a classifier or selection method doesn't accept the outcome, e.g. "CoxPH can't be used with classes". Before, every cross-validation failed and the error was reported afterwards.
• R/interfaceTemplate.R states the interface contract (arguments, prediction formats, feature extractors). The developer vignettes describe adding a method with a registry entry.
• test-registry.R fits and predicts every registered keyword for each outcome it accepts, and checks the prediction format. A new method is tested by adding its entry.

Evidence

• Same machinery as before: against 15/ Defaults and dependencies: no rebalancing and serial by default; five packages to Imports, three dropped #141's build, these are identical:

  • available()'s keywords and descriptions;
  • the table of display names;
  • the TrainParams/PredictParams of every classifier keyword with tuneParams = NULL, "auto" or a list;
  • the function of every selection keyword.

  (ProjectClassifyR scripts/structure/registry_evidence.R.)

• Regression harness: all 18 scenarios are identical to 15/ Defaults and dependencies: no rebalancing and serial by default; five packages to Imports, three dropped #141's build, including the characteristics recorded in results.

• 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, the undeclared global variables of the plotting code. Its test run passes 526 expectations, none failing.

🤖 Generated with Claude Code

https://claude.ai/code/session_01TfsjwkbDbHkX4kJZNe4ER4