Kernel Mean Matching

Description

Kernel Mean Matching (KMM) is an algorithm first proposed by Huang et al. [1] for instance weighting in domain adaptation. The algorithm finds weights associated with the source domain data points in such a way that the Maximum Mean Discrepancy of source and target domains is minimized in a reproducing kernel Hilbert space.

• Class
• Methods
  • fit
  • fit_predict
• Attributes
  • weights_

Class

class transfer.instance_weighting.KernelMeanMatching(B = 1000, epsilon = 0.1, kernel = "gaussian", gamma = None, coef0 = 1.0, degree = 3)

Parameter	Type	Description
B	float	Upper constraint on the weights. Acts as a form of regularization.
epsilon	float	Allowed distance from the mean of the weights to one.
kernel	string or callable	Kernel used. Can be a callable that takes two matrices of points and returns their kernel matrix or one of {"gaussian", "linear", "polynomial", "sigmoid", "laplacian"}.
gamma	float or None	Coefficient of the dot product of points in the kernel. If None it defaults to 0 / n_features.
coef0	float	The coefficient term in the kernel when choosing "polynomial" or "sigmoid". It is ignored in other cases.
degree	int	The degree of the kernel when choosing "polynomial". It is ignored in other cases.

Methods

fit

fit(Xs, Xt)

Fits the model to the given source and target domains' inputs by computing the source domain weights.

Parameter	Type	Description
Xs	array of shape (n_samples_s, n_features)	Source domain inputs.
Xt	array of shape (n_samples_t, n_features)	Targrt domain inputs.

Returns: self, fitted weighter.

fit_predict

fit_predict(Xs, Xt)

Fits the model to the given source and target domains' inputs and returns the computed weights.

Parameter	Type	Description
Xs	array of shape (n_samples_s, n_features)	Source domain inputs.
Xt	array of shape (n_samples_t, n_features)	Targrt domain inputs.

Returns: beta, array of shape (n_samples_s) containing the weights of the source domain's inputs.

Attributes

weights_

weights_

Array of shape (n_samples_s) containing the weights of the source domain's inputs.

References

1. J. Huang et al. Correcting sample selection bias by unlabeled data. 2006. https://papers.nips.cc/paper/2006/file/a2186aa7c086b46ad4e8bf81e2a3a19b-Paper.pdf