neural_trees.distillation.munge

neural_trees.distillation.munge(X, passes: int = 2, p: float = 0.5, s: float = 5.0, random_state=None) → ndarray

Synthetic rows near the rows of X, for a teacher to label (MUNGE, Bucila, Caruana and Niculescu-Mizil, “Model compression”, KDD 2006).

Each pass makes one new row per row of X: feature by feature, with probability p, the value is replaced by one drawn around the row’s nearest neighbour’s (normal, sd = abs(difference) / s); a feature with at most ten distinct values copies the neighbour’s value instead. Neighbours are found on standardised features, so the scale of X does not matter. A missing value stays missing unless it is replaced by the neighbour’s.

Returns an array of shape (passes * len(X), n_features).