Fix multiclass SVC with non-uniform class weights

[viclafargue]

Fixes #8578.

Multiclass SVC converted class weights into per-sample weights but did not pass them to the binary classifiers. This caused incorrect predictions with non-uniform class weights.

This PR explicitly forwards the correct weights:

• One-vs-one classifiers receive weights for their selected class pair.
• One-vs-rest classifiers receive the complete weight array.
• Binary classifiers no longer reinterpret the original class-weight mapping.

Regression tests cover class weights, sample weights, dense and sparse inputs, numeric and string labels, and both multiclass strategies.

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[viclafargue]

Following Scikit-Learn's logic over here.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: 6eef85b2-d38e-4589-8ce4-9884fc4cd67a

📥 Commits

Reviewing files that changed from the base of the PR and between 1c3db4b and a274f69.

📒 Files selected for processing (3)

• python/cuml/cuml/multiclass/multiclass.py
• python/cuml/cuml/svm/svc.py
• python/cuml/tests/test_svm.py

Included review availability: Your plan provides up to 12 included reviews per hour; 11 remain after this review.

---

📝 Summary

Summary by CodeRabbit

• New Features

  • Multiclass SVC models now support per-sample weighting during training.
  • Weighted One-vs-One and One-vs-Rest classification is supported for dense and sparse inputs.
  • Class labels can be numeric or string values when using custom or balanced class weights.

• Bug Fixes

  • Improved handling of class and sample weights to prevent weights from being applied more than once.

• Documentation

  • Updated multiclass classifier documentation to describe weighted fitting behavior.

Walkthrough

The change adds weighted One-vs-One and One-vs-Rest fitting, propagates weights through multiclass SVC, prevents class-weight duplication, preserves feature metadata, and adds comparisons with scikit-learn.

Changes

Weighted multiclass SVC

Layer / File(s)	Summary
Weighted multiclass fitting python/cuml/cuml/multiclass/multiclass.py	_BaseMulticlassClassifier.fit accepts sample_weight. Weighted OvO and OvR fitting trains cloned binary estimators with validated weights and preserves feature metadata. Documentation describes the cuML fitting path.
SVC weight wiring python/cuml/cuml/svm/svc.py	Multiclass SVC passes processed sample weights to the wrapper and clears binary class_weight after class weights are incorporated into the sample weights.
Multiclass SVC validation python/cuml/tests/test_svm.py	Tests cover custom, per-sample, and balanced class weights across dense and sparse inputs, label types, and OvO/OvR decision functions.

Priority: ➖ Normal

Estimated code review effort: 4 (Complex) | ~45 minutes

Change: Bug fix · Severity of issue fixed: Medium

Merge Risk: ⚪ Minimal · up to e3dd1

Weighted multiclass SVC fitting now propagates weights through the binary estimators and is validated against scikit-learn across the stated strategy and input variants. No merge-blocking risk remains.

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 30.00% which is insufficient. The required threshold is 80.00%. Docstring coverage is scoped to functions touched by this diff. Analyzed 10 functions across 3 files.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly identifies the main change: fixing multiclass SVC behavior with non-uniform class weights.
Description check	✅ Passed	The description directly explains the class-weight propagation issue, the implemented fix, and the regression test coverage.
Linked Issues check	✅ Passed	The change meets the coding requirements in issue #8578. SVC._fit_multiclass forwards processed sample weights and disables binary class_weight reuse. The weighted OvO helper selects each class pa…
Out of Scope Changes check	✅ Passed	The changes stay within issue #8578. The generic multiclass fitting helpers are required to propagate pair-specific and complete weights. The sample_weight support and documentation updates support …

✨ Finishing Touches 💡 1

🛠️ Fix failing CI checks 💡

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

_{Comment @coderabbitai help to get the list of available commands.}