Optimize a multiclass problem

QuOptuna handles binary and multiclass targets with different conventions:

	Binary	Multiclass
Target encoding	`{-1, +1}`	codes `0..K-1`
Objective metric	binary F1	macro-F1
Variational quantum models	native	automatically OvR-wrapped (one-vs-rest, K binary sub-models)
Kernel / classical models	native	native

Variational quantum models are automatically wrapped in one-vs-rest (OvR): K binary sub-models, one per class. Kernel and classical sklearn models handle multiclass natively without wrapping.

Run a 3-class problem (CLI)

Iris (--uci-id 53) has 3 classes and is a good smoke test:

quoptuna optimize --uci-id 53

QuOptuna detects the number of classes, switches the objective to macro-F1, and OvR-wraps any variational quantum models in the search space.

Python path

From Python, pass the raw target column — QuOptuna reads the class codes directly:

from quoptuna.backend.tuners.optimizer import Optimizer

Optimizer(
    study_name="iris-3class",
    db_name="iris",
    data=data,            # raw target column with codes 0..K-1
    model_types=models,
    search_space=space,
).optimize(n_trials=100)

Speed up OvR sub-fits

OvR sub-fits are serial by default. To fit the K binary sub-models concurrently on threads, set the environment variable:

export QUOPTUNA_OVR_N_JOBS=3   # = K, the number of classes

Threads only — QuOptuna never uses processes here because pickling JAX models is unsafe.

Multiclass fairness

For fairness-aware multiclass search you must set favorable_class (which class code counts as the favorable outcome). See Run fairness-aware search.