- recpack.algorithms.loss_functions.warp_loss_wrapper(X_true: scipy.sparse._csr.csr_matrix, X_pred: scipy.sparse._csr.csr_matrix, batch_size: int = 1000, num_negatives: int = 20, margin: float = 1.9, sample_size=None, exact=False)
Metric wrapper around the
Positives and negatives are sampled from the X_true matrix using
recpack.algorithms.samplers.WarpSampler. Their scores are fetched from the X_pred matrix.
X_true (csr_matrix) – True interactions expected for the users
X_pred (csr_matrix) – Predicted scores.
batch_size (int, optional) – Size of the sample batches, defaults to 1000
num_negatives (int, optional) – How many negatives to sample for each positive item, defaults to 20
margin (float, optional) – required margin between positives and negatives, defaults to 1.9
sample_size (int, optional) – How many samples to construct
exact (bool, optional) – If True sampling happens exact, otherwise sampling assumes high sparsity of data, accepting a minimal amount of false negatives. This speeds up sampling without significant loss of quality, defaults to False
The warp loss
- Return type