recpack.algorithms.loss_functions.covariance_loss

recpack.algorithms.loss_functions.covariance_loss(H: torch.nn.modules.sparse.Embedding, W: torch.nn.modules.sparse.Embedding) → torch.Tensor

Covariance loss.

Convariance loss as described in Cheng-Kang Hsieh et al., Collaborative Metric Learning. WWW2017 http://www.cs.cornell.edu/~ylongqi/paper/HsiehYCLBE17.pdf

The loss term is used to penalize covariance between embedding dimensions and thus disentangle these embedding dimensions.

It is assumed H and W are embeddings in the same space.

Parameters

• H (nn.Embedding) – Item embedding

• W (nn.Embedding) – User Embedding

Returns

Covariance loss term

Return type

torch.Tensor