Add Optuna-based hyperparameter optimization in gpredomicspy

[eprifti]

Context

gpredomics has many tunable parameters (k_penalty, population_size, cooling_rate, etc.) that significantly affect model quality. Manual tuning is tedious and suboptimal. Optuna provides efficient Bayesian hyperparameter optimization.

Design

Implement in gpredomicspy (Python layer) as gpredomicspy.optimize():

import gpredomicspy as gp

# Define search space
search_space = {
    "algo": ["ga", "beam", "sa", "ils", "lasso"],
    "k_penalty": (1e-5, 0.01, "log"),
    "language": ["ter", "bin,ter", "bin,ter,ratio"],
    "data_type": ["prev", "raw", "raw,prev"],
    "population_size": (500, 10000),
    "cooling_rate": (0.99, 0.9999),
    "feature_minimal_prevalence_pct": (5, 30),
    "feature_maximal_adj_pvalue": (0.01, 0.1),
}

results = gp.optimize(
    base_param="param.yaml",
    search_space=search_space,
    n_trials=100,
    metric="test_auc",        # or "fit", "spearman", etc.
    direction="maximize",
    cv=True,                  # use CV for robust evaluation
    n_jobs=4,                 # parallel trials
)

print(results.best_params)
print(results.best_value)
results.plot_importance()     # which params matter most

Parameters worth optimizing

Category	Parameter	Type	Range
Regularization	k_penalty	log-float	[1e-5, 0.1]
Regularization	fr_penalty	float	[0, 1]
Regularization	bias_penalty	float	[0, 1]
Algorithm	algo	categorical	ga/beam/sa/ils/lasso/aco
Data	language	categorical	ter/bin/ratio combos
Data	data_type	categorical	raw/prev/log combos
Feature selection	prevalence_pct	float	[5, 50]
Feature selection	max_adj_pvalue	float	[0.01, 1.0]
GA	population_size	int	[500, 10000]
GA	max_epochs	int	[50, 500]
GA	mutated_children_pct	float	[50, 95]
SA	cooling_rate	float	[0.99, 0.9999]
SA	max_iterations	int	[1000, 50000]
ACO	n_ants	int	[50, 1000]
ACO	alpha/beta	float	[0.5, 5.0]
ACO	rho	float	[0.01, 0.5]

Web app integration

Add a "Tune" button in ParametersTab that:

1. Opens a modal with parameter search space configuration
2. Runs Optuna in the background (via worker.py)
3. Shows convergence plot + parameter importance
4. "Apply best" button to set optimal parameters

Dependencies

• optuna (pip install optuna)
• optuna-dashboard (optional, for web visualization)

[delahondes]

If linked to scitq (or more generally to large parallel executions of learnings), the idea would be to start with a grid approach and to opportunisticly "clean" the grid by invalidating certain tasks.