Why does calling FuzzyCMeans.predict_proba() alter the centroids of the fitted model?

[ensley-nexant]

I'm trying to understand why I can call .predict_proba() multiple times on the same fitted model, passing the same function, and get different predictions. Is this the intended behavior?

It looks like the centroids are being updated on every prediction, and I'm not sure why that would be the case. Here is an example:

from skfda import datasets
from skfda.ml.clustering import FuzzyCMeans

X, y = datasets.fetch_weather(return_X_y=True, as_frame=True)
fd = X.iloc[:, 0].values.coordinates[0]

kmeans = FuzzyCMeans(n_clusters=4, random_state=2)
kmeans.fit(fd)

print("Initial centroids")
print(kmeans.cluster_centers_.data_matrix[0][:4])

for i in range(1, 5):
    print("\nPrediction", i)
    print(kmeans.predict_proba(fd[1])[0])
    print("Centroids (head):")
    print(kmeans.cluster_centers_.data_matrix[0][:4])

Output:

Initial centroids
[[-7.48331886]
 [-7.69572714]
 [-8.33347371]
 [-8.2457692 ]]

Prediction 1
[0.81649987 0.04951254 0.12522585 0.00876174]
Centroids (head):
[[-4.4]
 [-4.2]
 [-5.3]
 [-5.4]]

Prediction 2
[0.2 0.2 0.2 0.4]
Centroids (head):
[[-4.4]
 [-4.2]
 [-5.3]
 [-5.4]]

Prediction 3
[0.28571429 0.28571429 0.28571429 0.14285714]
Centroids (head):
[[-4.4]
 [-4.2]
 [-5.3]
 [-5.4]]

Prediction 4
[0.25 0.25 0.25 0.25]
Centroids (head):
[[-4.4]
 [-4.2]
 [-5.3]
 [-5.4]]

[vnmabus]

Good catch! This is indeed NOT intended behavior. Calling predict or predict_proba should not change the model. I will create an issue from this discussion to track this bug.