Transforming new data

A transform should place new rows in an existing coordinate system. Fitting an embedding again on a new batch does not provide that guarantee.

Native transforms

PCA stores its fitted mean and components. Its Rust fit(X) returns None, so call it separately instead of chaining .fit(X).transform(...).

import numpy as np
from sklearn.datasets import load_digits
from squeeze import PCA

X = np.asarray(load_digits().data, dtype=np.float64)
model = PCA(n_components=2)
model.fit(X[:1000])
new_coordinates = model.transform(X[1000:])
assert new_coordinates.shape == (797, 2)

UMAP also supplies transform using the fitted model. Some modes have additional restrictions, including densMAP; do not assume every UMAP option supports inference. Other documented Rust reducers have no native out-of-sample transform.

Interpolation for fit-only methods

OutOfSampleDR fits the base reducer once, retains the training data, and interpolates new coordinates from nearest training neighbors. This is a wrapper approximation, not an inference method learned by NeighborMap or t-SNE.

import numpy as np
from sklearn.datasets import load_digits
from squeeze import NeighborMap, OutOfSampleDR

X = np.asarray(load_digits().data[:200], dtype=np.float64)
model = OutOfSampleDR(NeighborMap(random_state=42), n_neighbors=5, weights="distance")
training_coordinates = model.fit_transform(X[:150])
new_coordinates = model.transform(X[150:])
assert training_coordinates.shape == (150, 2)
assert new_coordinates.shape == (50, 2)

Weights may be "distance" or "uniform". Novel clusters outside the training support can interpolate poorly; evaluate on held-out data before using this route. UMAP's inverse_transform is likewise an approximation, not lossless reconstruction. Rust PCA does not expose an inverse_transform method.