KDERepresentation

class mlquantify.representations.KDERepresentation(bandwidth=0.1, kernel='gaussian')

Kernel density estimation representation.

This representation fits a kernel density estimator (KDE) to the training data for each class, using the specified bandwidth and kernel. The test-time representation is simply the raw feature vector, and per-class likelihoods can be obtained by evaluating the fitted KDEs on the test samples. This allows for a non-parametric density-based representation of the data, which can capture complex class distributions without assuming a specific parametric form.

Parameters:

bandwidthfloat, default=0.1

The bandwidth parameter for the KDE, controlling the smoothness of the density estimate. Smaller values lead to a more flexible fit that can capture finer details of the data distribution, while larger values produce a smoother estimate that may overlook small-scale structure.

kernelstr, default=”gaussian”

The kernel to use for the KDE. Options are: - “gaussian”: Gaussian kernel (the default), which produces a smooth density estimate. - “tophat”: Uniform kernel, which gives equal weight to all points within the bandwidth radius and zero weight to points outside. - “epanechnikov”: Epanechnikov kernel, which is a parabolic kernel that gives more weight to points closer to the center of the bandwidth and less weight to points farther away, with zero weight beyond the bandwidth radius. - “exponential”: Exponential kernel, which gives more weight to points closer to the center of the bandwidth and decays exponentially with distance, without a hard cutoff. - “linear”: Linear kernel, which gives weight that decreases linearly with distance from the center of the bandwidth, with zero weight beyond the bandwidth radius. - “cosine”: Cosine kernel, which gives weight that follows a cosine function of the distance from the center of the bandwidth, with zero weight beyond the bandwidth radius

Examples

>>> from mlquantify.representations._density import KDERepresentation
>>> import numpy as np
>>> rng = np.random.default_rng(0)
>>> X = rng.standard_normal((100, 2))
>>> y = (X[:, 0] > 0).astype(int)
>>> rep = KDERepresentation(bandwidth=0.2, kernel="gaussian").fit(X, y)
>>> rep.class_representations_[0]
KernelDensity(bandwidth=0.2, kernel='gaussian')

class_likelihoods(X)

Evaluate per-class kernel density likelihoods for test instances.

For each class KDE fitted during fit, scores every test sample and exponentiates the log-density to obtain raw likelihood values.

Parameters:

Xarray-like of shape (n_samples, n_features)

Test feature matrix.

Returns:

likelihoodsndarray of shape (n_classes, n_samples)

Per-class likelihood for each test instance, where likelihoods[c, i] is the density of class c at sample i.

Examples

>>> from mlquantify.representations._density import KDERepresentation
>>> import numpy as np
>>> rng = np.random.default_rng(0)
>>> X = rng.standard_normal((100, 2))
>>> y = (X[:, 0] > 0).astype(int)
>>> rep = KDERepresentation().fit(X, y)
>>> lkl = rep.class_likelihoods(X[:5])
>>> lkl.shape
(2, 5)

fit(X, y, classes=None, sample_weight=None)

Fit the representation to labelled training data.

Validates shapes, stores the class labels, delegates internal fitting to _fit, and verifies that the subclass set class_representations_ during that call.

Parameters:

Xarray-like of shape (n_samples, n_features) or (n_samples,)

Feature matrix or pre-computed score array for the training instances.

yarray-like of shape (n_samples,)

Class labels for each training instance.

classesarray-like of shape (n_classes,) or None, default=None

Explicit list of class labels. If None, the unique values in y are used.

sample_weightarray-like of shape (n_samples,) or None, default=None

Per-sample weights forwarded to _fit.

Returns:

selfBaseRepresentation

The fitted representation object.

Raises:

ValueError

If X and y have inconsistent lengths or X is zero-dimensional.

AttributeError

If the subclass did not define class_representations_ inside _fit.

Examples

>>> from mlquantify.representations import HistogramRepresentation
>>> import numpy as np
>>> X = np.random.default_rng(0).uniform(0, 1, (100, 1))
>>> y = (X[:, 0] > 0.5).astype(int)
>>> rep = HistogramRepresentation(bins=(5,)).fit(X, y)
>>> rep.class_representations_.shape
(2, 5)

transform(X)

Return the input as a float array (identity transform).

The KDE representation uses raw feature vectors as the test-time representation; density evaluation happens in class_likelihoods.

Parameters:

Xarray-like of shape (n_samples, n_features)

Test feature matrix.

Returns:

X_transformedndarray of shape (n_samples, n_features)

Input cast to float64.

Examples

>>> from mlquantify.representations._density import KDERepresentation
>>> import numpy as np
>>> rep = KDERepresentation()
>>> X = np.array([[0.1, 0.2], [0.3, 0.4]])
>>> rep.transform(X).shape
(2, 2)