lenskit.graphs

Graph-based models, especially GNNs with torch_geometric.

class lenskit.graphs.LightGCNConfig(*, embedding_size=64, layer_count=2, layer_blend=None, batch_size=8192, learning_rate=0.01, epochs=10, regularization=0.01, loss='pairwise')

Bases: EmbeddingSizeMixin, BaseModel

Configuration for LightGCNScorer.

Stability:

Experimental

Parameters:

• embedding_size (Annotated[int, Gt(gt=0)])

• layer_count (Annotated[int, Gt(gt=0)])

• layer_blend (Annotated[float, Gt(gt=0)] | list[Annotated[float, Gt(gt=0)]] | None)

• batch_size (Annotated[int, Gt(gt=0)])

• learning_rate (Annotated[float, Gt(gt=0)])

• epochs (Annotated[int, Gt(gt=0)])

• regularization (Annotated[float, Gt(gt=0)] | None)

• loss (Literal['logistic', 'pairwise'])

model_config: ClassVar[ConfigDict] = {}

Configuration for the model, should be a dictionary conforming to [ConfigDict][pydantic.config.ConfigDict].

embedding_size: PositiveInt

The dimension of the embedding space (number of latent features). Seems to work best as a power of 2.

layer_count: PositiveInt

The number of layers to use.

layer_blend: PositiveFloat | list[PositiveFloat] | None

The blending coefficient(s) for layer blending. This is equivalent to alpha in LightGCN.

batch_size: PositiveInt

The training batch size.

learning_rate: PositiveFloat

The learning rate for training.

epochs: PositiveInt

The number of training epochs.

regularization: PositiveFloat | None

The regularization strength.

loss: Literal['logistic', 'pairwise']

The loss to use for model training.

pairwise

BPR pairwise ranking loss, using LightGCN.recommend_loss().

logistic

Logistic link prediction loss, using LightGCN.link_pred_loss().

class lenskit.graphs.LightGCNScorer(config=None, **kwargs)

Bases: UsesTrainer, Component[ItemList, …]

Scorer using LightGCN [].

Stability:

Experimental

Parameters:

• config (LightGCNConfig)

• kwargs (Any)

create_trainer(data, options)

Create a model trainer to train this model.

to(device)

Move the model to a different device.

Modules

lightgcn

LightGCN recommendation.