BayesianGaussianMixture fit takes 1000 times as long on a subset - why?

[russell-at-qudo]

I'm using BayesianGaussianMixture to create some segments of data, and it works really well. The issue I am having is that I want to test relabel consistency, and I am doing this by creating a stratified train/test split (stratified by segment label). When I am doing this, the fit process is taking around 1000 times as long to fit the training set than the full dataset given the same model parameters, and I don't know why!?

model = BayesianGaussianMixture(n_components=5, n_init=42)
model.fit(data)

This takes around 1.2 seconds for my dataset.

X_train, X_test, y_train, y_test = train_test_split(data, cluster_labels, test_size=0.1,
                                                            random_state=42,
                                                            stratify=cluster_labels)
model = BayesianGaussianMixture(n_components=5, n_init=42)
model.fit(X_train)

This takes about 1400 seconds.

Any ideas?