Include a lower bound attribute of BaseMixture

[a-wozniakowski]

Describe the workflow you want to enable

Currently, there exists a lower_bound_ attribute in the fit_predict method of BaseMixture. However, the entire sequence of lower bounds is not accessible, which makes a convergence analysis more difficult to a user.

Describe your proposed solution

In addition to the lower_bound_ attribute, create a new attribute called lower_bounds_, which is a list of floats where each float is a lower bound set in

scikit-learn/sklearn/mixture/_base.py

Line 248 in 92c9b18

lower_bound = self._compute_lower_bound(log_resp, log_prob_norm)

.

Describe alternatives you've considered, if relevant

No response

Additional context

Creating the list and appending to it would increase memory costs (but not by much). Is this a possible concern?

[myenugula]

Sounds like a nice addition! I'll work on modifying the code to add this feature with some samples.

[myenugula]

@a-wozniakowski I have added the lower bounds list feature here. Please give it a shot and let me know. I'll make a PR if it's all good.

Here's a code sample for using it on Iris dataset:

from sklearn import datasets
from sklearn.mixture import GaussianMixture
import matplotlib.pyplot as plt
import pandas as pd

iris = datasets.load_iris()
X = iris.data

d = pd.DataFrame(X)

gmm = GaussianMixture(n_components=3, n_init=10)
gmm.fit(d)

lower_bounds = gmm.lower_bounds_
n_iter = gmm.n_iter_

plt.figure(figsize=(10, 6))
plt.plot(range(1, n_iter + 1), lower_bounds, marker='o', linestyle='-')
plt.title('GMM Lower Bounds Across Iterations')
plt.xlabel('Iteration')
plt.ylabel('Lower Bound')
plt.grid(True)
plt.show()

[glemaitre]

Yep we could have this feature.