Google I/O is a wrap! Catch up on TensorFlow sessions View sessions

tfrs.layers.feature_interaction.DotInteraction

Dot interaction layer.

tfrs.layers.feature_interaction.DotInteraction(
    self_interaction: bool = False,
    skip_gather: bool = False,
    name: Optional[str] = None,
    **kwargs
) -> None

See theory in the DLRM paper: https://arxiv.org/pdf/1906.00091.pdf, section 2.1.3. Sparse activations and dense activations are combined. Dot interaction is applied to a batch of input Tensors [e1,...,e_k] of the same dimension and the output is a batch of Tensors with all distinct pairwise dot products of the form dot(e_i, e_j) for i <= j if self self_interaction is True, otherwise dot(e_i, e_j) i < j.

self_interaction Boolean indicating if features should self-interact. If it is True, then the diagonal enteries of the interaction matric are also taken.
skip_gather An optimization flag. If it's set then the upper triangle part of the dot interaction matrix dot(e_i, e_j) is set to 0. The resulting activations will be of dimension [num_features * num_features] from which half will be zeros. Otherwise activations will be only lower triangle part of the interaction matrix. The later saves space but is much slower.
name String name of the layer.

Methods

call

View source

call(
    inputs: List[tf.Tensor]
) -> tf.Tensor

Performs the interaction operation on the tensors in the list.

The tensors represent as transformed dense features and embedded categorical features. Pre-condition: The tensors should all have the same shape.

Args
inputs List of features with shapes [batch_size, feature_dim].

Returns
activations Tensor representing interacted features. It has a dimension num_features * num_features if skip_gather is True, otherside num_features * (num_features + 1) / 2 if self_interaction is True and num_features * (num_features - 1) / 2 if self_interaction is False.