Datasets¶

cca_zoo.datasets provides utilities for generating synthetic multiview data and loading small real-world datasets.

`make_joint_data` — simulated multiview data¶

make_joint_data generates data from a linear latent variable model, in the style of sklearn.datasets' make_* generators:

\[ X_i = Z W_i^\top + E_i \]

where:

\(Z \in \mathbb{R}^{n \times k}\) is the shared latent variable (\(k\) = n_components)
\(W_i \in \mathbb{R}^{p_i \times k}\) is the view-specific loading matrix
\(E_i\) is independent Gaussian noise, with variance controlled by signal_to_noise

For independent training and test sets, generate once and split every view together with sklearn.model_selection.train_test_split:

from sklearn.model_selection import train_test_split

from cca_zoo.datasets import make_joint_data

views = make_joint_data(
    n_samples=400,
    n_features=[50, 40],  # different feature counts per view
    n_components=2,
    signal_to_noise=2.0,  # higher = less noise
    random_state=0,
)
X1_train, X1_test, X2_train, X2_test = train_test_split(*views, random_state=0)

Parameters¶

Parameter	Type	Default	Description
`n_samples`	`int`	`100`	Observations
`n_features`	`int` or `list[int]`	`10`	Features per view (scalar broadcasts)
`n_views`	`int`	`2`	Number of views
`n_components`	`int`	`1`	Shared latent dimension
`signal_to_noise`	`float` or `list[float]`	`1.0`	SNR per view (higher = less noise)
`random_state`	`int`, `RandomState` or `None`	`None`	Seed for reproducibility
`return_latent`	`bool`	`False`	Also return the latent variable `z`

Usage patterns¶

# Three views, same feature count (scalar broadcasts)
views = make_joint_data(n_samples=100, n_features=20, n_views=3, n_components=2)

# Different SNR per view
views = make_joint_data(n_samples=100, n_features=20, signal_to_noise=[4.0, 1.0])

# The true latent variable too, e.g. to check recovery
views, z = make_joint_data(n_samples=100, n_components=2, return_latent=True)

Toy real-world datasets¶

Two small real-world datasets are included for quick experimentation:

`load_linnerud`¶

Wraps sklearn.datasets.load_linnerud. Like sklearn's loaders it returns a Bunch, with the views in views and each view's feature_names; return_views=True returns the views alone, as sklearn's return_X_y:

View 1: exercise measurements (chin-ups, sit-ups, jumps) — shape (20, 3)
View 2: physiological measurements (weight, waist, pulse) — shape (20, 3)

from cca_zoo.datasets import load_linnerud

exercise, physiological = load_linnerud(return_views=True)
print(exercise.shape)  # (20, 3)
print(physiological.shape)  # (20, 3)

`load_breast_cancer`¶

Wraps sklearn.datasets.load_breast_cancer. Splits the 30 features into two halves to create a two-view dataset:

View 1: first 15 features — shape (569, 15)
View 2: last 15 features — shape (569, 15)

from cca_zoo.datasets import load_breast_cancer

view1, view2 = load_breast_cancer(return_views=True)
print(view1.shape)  # (569, 15)
print(view2.shape)  # (569, 15)