[ellispatrick]

Part 15 (#140 → this → #142 → #143). Based on #140, which holds #135–#139 together. Each commit can be accepted or rejected on its own.

Summary

crossValidate results are unchanged; defaults of runTest and runTests change.

• No class rebalancing by default (Q33): ModellingParams() no longer downsamples by default.
• Serial by default (Q34): CrossValParams() runs on one core by default.
• Fewer attached packages (Q37): generics, S4Vectors, MultiAssayExperiment, BiocParallel and survival move from Depends to Imports, and ggpubr, reshape2 and broom are dropped.
• easyHard fix: its lambda filter no longer drops a lambda whose coefficients sum to zero.

Why

Ellis Patrick's decisions, ProjectClassifyR docs/DECISIONS.md, 10 October 2026:

• Q33, "Default 'none'":
  • runTests and crossValidate never applied the downsampling default; only a direct runTest call did (audit Q-C1).
  • The default now matches what cross-validation does.
• Q34, "Serial":
  • bpparam() gave as many workers as the computer has cores, less two (125 on our server), and no seed.
  • So default runTests runs were not reproducible.
  • Serial runs are reproducible after set.seed.
• Q37, "Both":
  • Bioconductor's guidance is to import packages rather than attach them.
  • The three dropped packages were each used once.

What changes

• ModellingParams(balancing = c("none", "downsample", "upsample")):
  • The help page and vignette say the default is no rebalancing.
  • Direct runTest calls no longer downsample the training samples unless asked.
• CrossValParams(parallelParams = SerialParam()):
  • The help page explains how to use several cores reproducibly (e.g. MulticoreParam(workers = 4, RNGseed = 1)).
  • crossValidate is unaffected; it already built its own seeded parameter.
• Depends → Imports:
  • library(ClassifyR) attaches 1 package instead of 15, and puts 52 objects on the search path instead of 1,602.
  • train, Surv and DataFrame are re-exported, so code using them keeps working.
  • Users who call functions such as SerialParam, MulticoreParam, MultiAssayExperiment, colData, mcols or Pairs directly now need library() of that package, or pkg::.
  • The examples and the vignette now qualify such calls.
• Dropped packages:
  • ggpubr: crissCrossPlot draws its two heatmaps (feature selection and random features) as two facets of one ggplot with a shared legend. Before, ggarrange combined two plots.
  • reshape2: the long table is built with base R.
  • broom: easyHard makes the coefficient table (term, estimate, std.error, statistic, p.value) from summary(glm). The "full" mode returns a DataFrame, as documented; it returned a tibble.
• easyHard(fitMode = "full"): the lambda filter keeps lambdas at which any coefficient is not zero. The signed sum could be zero by cancellation.

Evidence

• Tests and check:
  • R CMD check's test run passes (439 expectations, none failing, 2 skipped).
  • R CMD check (tests, examples, vignettes) gives 1 NOTE, the existing undeclared global variables of the plotting code.
• Regression harness: all 18 scenarios are identical to int/ #135–#139 together for testing (not for merging) #140's build.
• easyHard: single and full modes on asthma with clinical covariates are identical to the broom version.
• crissCrossPlot:
  • With one panel (modelTrain), the plot is identical to before.
  • With two panels (modelTest with random features), it shows the same tiles, values and diagonal outlines as before, in two facets with one legend.
• Scripts and images: ProjectClassifyR scripts/structure/deps_evidence.R, results/structure/.

🤖 Generated with Claude Code

https://claude.ai/code/session_01TfsjwkbDbHkX4kJZNe4ER4