wwu-mmll / confound_corrected_cpm / 29932827040

86%

main: 85%

push

github

Add presence filter, fix GPU + RepeatedKFold bugs, remove dead code

Features
- Edge selection: new `presence_filter` (bool/float) on UnivariateEdgeSelection
  keeps only edges nonzero in >= a fraction of subjects, for dropping structural
  zeros in sparse connectomes (e.g. DTI). Computed per fold on training subjects
  from X only (no target leakage), additive to the variance gate. Replaces the
  never-functional `t_test_filter` (now deprecated: warns and is ignored).

Fixes
- GPU device mismatch in confound-controlled edge selection: get_residuals built
  the intercept column on the CPU, crashing the `*_partial` statistics on
  device='cuda'. Intercept now created on the inputs' device; numpy-return path
  goes through .cpu(). Added a CUDA-guarded regression test.
- RepeatedKFold individual-level outputs: predictions/network strengths are now
  tagged with a `repeat` id (network strengths also with `sample_index`), and the
  HTML report averages each subject across repeats before plotting. Previously
  every subject appeared n_repeats times in the scatter and pooled-r annotation.
  Edge stability and the fold-metric summary still pool all folds+repeats.

Cleanup (no behavior change)
- Removed dead/broken utils converters (vector_to_matrix_3d and two triangular
  helpers) + a duplicate import; ResultsManager.collect_results (undefined names)
  and unused _save_inner_cv_to_csv; SelectPercentile/SelectKBest stubs; the
  duplicated PThreshold docstring; the unused simulation/simulate_multivariate
  module; and AI-narration comments in linear_model.py. cv_predictions.csv is now
  written once (was written twice per run).

Docs: methods.md (presence filter), interpreting_results.md (repeat column),
CHANGELOG (Unreleased), RELEASE_PLAN (refreshed + reprioritized).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

44 of 46 new or added lines in 5 files covered. (95.65%)

2306 of 2683 relevant lines covered (85.95%)

0.86 hits per line