GMM test

[Haffi112]

This is just a sanity check to see if GMMs produce more stable results than the kernel density estimator. The aim is not necessarily to merge this with main, but just to check if it produces better results.

[Copilot]

Pull Request Overview

This PR replaces the kernel density estimator with a Gaussian Mixture Model (GMM) for generative modeling of EU survey data. The change aims to evaluate whether GMMs provide more stable results than the previous approach.

• Replaces KernelDensity with GaussianMixture from sklearn
• Adds component selection methodology with configurable parameters for GMM optimization
• Introduces validation-based model selection using BIC/AIC criteria

Reviewed Changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 3 comments.

File	Description
src/scripts/train_generative_model.py	Passes new GMM configuration parameters to the training function
src/european_values/generative_training.py	Core implementation change from KernelDensity to GaussianMixture with component selection
config/config.yaml	Adds GMM-specific configuration parameters with default values

---

_{Tip: Customize your code reviews with copilot-instructions.md. Create the file or learn how to get started.}

[Copilot]

The import of gmm_component_selection module will fail as this module is not included in the diff and likely doesn't exist yet. The functions evaluate_gmm_components, plot_component_selection, and save_evaluation_results are being imported but their implementation is missing.

Suggested change

	# --- Implementations for missing functions ---
	import numpy as np
	import matplotlib.pyplot as plt
	def evaluate_gmm_components(
	X_train,
	X_val,
	max_components,
	criterion="bic",
	covariance_type="full",
	random_state=None,
	n_init=1,
	reg_covar=1e-6,
	):
	"""Evaluate GMMs with different numbers of components and select the best one."""
	scores = []
	n_components_range = range(1, max_components + 1)
	for n in n_components_range:
	gmm = GaussianMixture(
	n_components=n,
	covariance_type=covariance_type,
	random_state=random_state,
	n_init=n_init,
	reg_covar=reg_covar,
	)
	gmm.fit(X_train)
	if criterion == "bic":
	score = gmm.bic(X_val)
	else:
	score = gmm.aic(X_val)
	scores.append(score)
	optimal_n = n_components_range[np.argmin(scores)]
	evaluation_results = {
	"n_components": list(n_components_range),
	"scores": scores,
	"criterion": criterion,
	}
	return optimal_n, evaluation_results
	def plot_component_selection(evaluation_results, output_path=None):
	"""Plot the selection criterion scores for different numbers of GMM components."""
	plt.figure()
	plt.plot(
	evaluation_results["n_components"],
	evaluation_results["scores"],
	marker="o",
	)
	plt.xlabel("Number of GMM components")
	plt.ylabel(evaluation_results.get("criterion", "Score").upper())
	plt.title("GMM Component Selection")
	if output_path:
	plt.savefig(output_path)
	else:
	plt.show()
	plt.close()
	def save_evaluation_results(evaluation_results, output_path):
	"""Save evaluation results to a CSV file."""
	df = pd.DataFrame({
	"n_components": evaluation_results["n_components"],
	"score": evaluation_results["scores"],
	})
	df.to_csv(output_path, index=False)

[Copilot]

The default value for reg_covar in the function signature (1e-6) differs from the config file value (1e-5). This inconsistency could lead to confusion about which value is actually used.

Suggested change

	reg_covar: 1e-5 # Regularization for covariance matrices (increased for extra stability)
	reg_covar: 1e-6 # Regularization for covariance matrices (matches function default)

[Copilot]

The default value for covariance_type in the function signature is 'full' but the config file sets it to 'diag'. This mismatch could cause unexpected behavior when the function is called without explicit parameters.

Suggested change

	covariance_type: str = "full",
	covariance_type: str = "diag",