polyview

polyview is a Python library for multi-view learning, providing a consistent API for multi-view clustering, embedding, and fusion methods. It allows you to easily build pipelines that combine native multi-view estimators with single-view sklearn tools, while keeping track of per-view data and predictions.

Documentation is available at: https://gwendal-debaussart.github.io/polyview/

Installation

Install from PyPI:

pip install polyview

Quick start

The snippet below demonstrates four common workflows:

• native multi-view clustering,
• early fusion into single-view clustering,
• per-view sklearn models followed by late fusion,
• imputation of missing views.

import numpy as np
import polyview as pv
from polyview.dataset.make_multiview_gaussian import make_multiview_gaussian
from sklearn.cluster import KMeans as SKKMeans
from sklearn.preprocessing import StandardScaler

# Synthetic multiview dataset
mvd = make_multiview_gaussian(
    n_samples=200,
    n_features=20,
    n_views=3,
    centers=4,
    random_state=0,
)
mvd.view_names = ["audio", "vision", "imu"]

# 1) Native multiview clustering
mv_model = pv.cluster.MultiViewKMeans(n_clusters=4, random_state=0)
mv_labels = mv_model.fit_predict(mvd)
print("MV labels:", mv_labels.shape)

# 2) Early fusion + single-view estimator
X_fused = pv.fusion.NormalizedFusion().fit_transform(mvd)
sv_model = SKKMeans(n_clusters=4, random_state=0)
sv_labels = sv_model.fit_predict(X_fused)
print("Fused shape:", X_fused.shape)

# 3) Per-view sklearn model + late fusion
# PolyPipeline runs one cloned SK estimator per view when final step is single-view.
pipe = pv.PolyPipeline(
    steps=[
        ("scale", StandardScaler()),
        ("kmeans", SKKMeans(n_clusters=4, random_state=0)),
    ]
)
labels_by_view = pipe.fit_predict(mvd.views)  # list of label vectors

late = pv.fusion.MajorityVote(weights=[0.2, 0.5, 0.3], tie_break="first")
late_labels = late.fit_predict(labels_by_view)
print("Late-fused labels:", late_labels.shape)

# 4) Missing views
# Views are incomplete when a modality was never acquired for a sample;
# missing entries are encoded as NaN.
incomplete, mask = pv.simulate_missing_views(mvd, missing_rate=0.2, random_state=0)
print("Missing rate per view:", pv.missing_rate(incomplete))

completed = pv.KNNViewImputer(n_neighbors=5).fit_transform(incomplete)
imputed_labels = pv.cluster.MultiViewKMeans(n_clusters=4, random_state=0).fit_predict(completed)
print("Labels from imputed views:", imputed_labels.shape)

Additional examples are available in the documentation: https://gwendal-debaussart.github.io/polyview/examples/

Contributing

Contributions are very welcome! If you plan to contribute, please see the CONTRIBUTING guidelines.

⚖️ License

GNU General Public License. See LICENSE for details.