adapt.metrics.neg_j_score

adapt.metrics.neg_j_score(Xs, Xt, max_centers=100, sigma=None)[source]

Compute the negative J-score between Xs and Xt.

\[\Delta = -\int_{\mathcal{X}} P(X_T) \log(P(X_T) / P(X_S))\]

Where:

\(P(X_S), P(X_T)\) are the probability density functions of Xs and Xt.

The source and target probability density functions are approximated with a mixture of gaussian kernels of bandwith sigma and centered in max_centers random points of Xt. The coefficient of the mixture are determined by solving a convex optimization (see [1])

Parameters

Xsarray: Source array
Xtarray: Target array
max_centersint (default=100): Maximum number of centers from Xt
sigmafloat (default=None): Kernel bandwidth. If None, the mean of pairwise distances between data from Xt is used.

Returns

scorefloat