DistMult

class DistMult(*, embedding_dim: int = 50, entity_initializer: str | ~typing.Callable[[~torch.Tensor], ~torch.Tensor] | None = <function xavier_uniform_>, entity_constrainer: str | ~typing.Callable[[~torch.Tensor], ~torch.Tensor] | None = <function normalize>, relation_initializer: str | ~typing.Callable[[~torch.Tensor], ~torch.Tensor] | None = <pykeen.utils.compose object>, regularizer: str | ~pykeen.regularizers.Regularizer | type[~pykeen.regularizers.Regularizer] | None = <class 'pykeen.regularizers.LpRegularizer'>, regularizer_kwargs: ~collections.abc.Mapping[str, ~typing.Any] | None = None, entity_representations_kwargs: ~collections.abc.Mapping[str, ~typing.Any] | None = None, relation_representations_kwargs: ~collections.abc.Mapping[str, ~typing.Any] | None = None, **kwargs)[source]

Bases: ERModel[Tensor, Tensor, Tensor]

An implementation of DistMult from [yang2014].

In this work, both entities and relations are represented by \(d\)-dimensional vectors stored in an Embedding matrix. The entity representation vectors are further constrained to have unit \(L_2\) norm. For the relation representations, a (soft) regularization term on the vector \(L_2\) norm is used instead.

The representations are then passed to the DistMultInteraction function to obtain scores.

This DistMult model can be seen as a simplification of the RESCAL model, where the relation matrices are restricted to diagonal matrices: Because of its restriction to diagonal matrices, DistMult is computationally cheaper than RESCAL, but at the same time it is less expressive. For example, it is not able to model anti-symmetric relations.

See also

Initialize DistMult.

Parameters:

  • embedding_dim (int) – The entity embedding dimension \(d\). Is usually \(d \in [50, 300]\).

  • entity_initializer (str | Callable[[Tensor], Tensor] | None) – The method used to initialize the entity embedding. Defaults to Xavier/Glorot uniform, c.f. OpenKE

  • entity_constrainer (str | Callable[[Tensor], Tensor] | None) – The constrainer for entity embeddings. Defaults to unit L2 norm.

  • relation_initializer (str | Callable[[Tensor], Tensor] | None) – The method used to initialize the relation embedding. Defaults to using Xavier/Glorot uniform first and then normalizing te unit L2 length.

  • regularizer (str | Regularizer | type[Regularizer] | None) – The relation representation regularizer.

  • regularizer_kwargs (Mapping[str, Any] | None) – Additional keyword-based parameters. Defaults to DistMult.regularizer_default_kwargs for the default regularizer.

  • entity_representations_kwargs (Mapping[str, Any] | None) – Additional parameters to entity_representations_kwargs passed to pykeen.models.ERModel. Note that those take precedence of those which are filled in by this class.

  • relation_representations_kwargs (Mapping[str, Any] | None) – Additional parameters to relation_representations_kwargs passed to pykeen.models.ERModel. Note that those take precedence of those which are filled in by this class.

  • kwargs – Remaining keyword arguments to forward to pykeen.models.ERModel

Note

The parameter pair (regularizer, regularizer_kwargs) is used for pykeen.regularizers.regularizer_resolver

An explanation of resolvers and how to use them is given in https://class-resolver.readthedocs.io/en/latest/.

Attributes Summary

hpo_default

The default strategy for optimizing the model's hyper-parameters

regularizer_default_kwargs

The LP settings used by [yang2014] for DistMult

Attributes Documentation

hpo_default: ClassVar[Mapping[str, Any]] = {'embedding_dim': {'high': 256, 'low': 16, 'q': 16, 'type': <class 'int'>}}

The default strategy for optimizing the model’s hyper-parameters

regularizer_default_kwargs: ClassVar[Mapping[str, Any]] = {'normalize': True, 'p': 2.0, 'weight': 0.1}

The LP settings used by [yang2014] for DistMult