Add multiclass classification (OVO, OVA)

[eprifti]

Context

gpredomics currently supports binary classification only (2 classes: 0 vs 1). Many clinical and biological problems involve multiple classes (e.g., disease subtypes, treatment response categories, multiple conditions).

Proposed approaches

One-vs-All (OVA / OVR)

• Train K binary classifiers, each separating one class from all others
• At prediction time, assign the class with the highest score/confidence
• Pros: simple, only K models needed, each model is a standard gpredomics binary model (fully interpretable)
• Cons: class imbalance (one class vs all others); models are not calibrated against each other; may produce ambiguous regions where multiple classifiers predict positive
• Implementation: can be orchestrated externally (run gpredomics K times with relabeled y), or integrated into the engine for convenience

One-vs-One (OVO)

• Train K×(K-1)/2 binary classifiers, one for each pair of classes
• At prediction time, each classifier votes; assign the class with the most votes
• Pros: each pairwise classifier sees balanced sub-problems; often better separation
• Cons: quadratic number of models; voting ties possible; harder to interpret the ensemble
• Implementation: more complex orchestration; needs a voting/aggregation layer

Comparison

Approach	# Models	Balance	Interpretability	Complexity
OVA	K	Imbalanced	High (each model is standalone)	Low
OVO	K(K-1)/2	Balanced	Medium (ensemble of pairwise models)	Medium

Design considerations

• Jury integration: The existing voting/jury system could potentially be reused for combining OVO classifiers
• Feature importance: How to aggregate feature importance across multiple binary models
• Cross-validation: CV should maintain class proportions across all classes (stratified K-fold)
• param.yaml: Need a new parameter for multiclass strategy (multiclass: ova or multiclass: ovo)
• Output format: Results should show per-class metrics (sensitivity, specificity, etc.) and a confusion matrix

Related work

• predomicsmc — existing multiclass extension via Predomics (R implementation)
• scikit-learn's OneVsRestClassifier / OneVsOneClassifier as design reference

Suggested implementation path

1. Start with OVA — simpler, each sub-model is a standard gpredomics run
2. Add OVO as an option later
3. Consider whether multiclass should be a core engine feature or an orchestration layer (wrapper script / predomicsapp-web feature)