cuq_graph_neural_network¶
- probly.method.cuq_graph_neural_network(input_encoder: Predictor[In, Out], latent_dim: int, num_classes: int, *, encoder_dim: int | None = None, num_flows: int = 6, class_counts: list | None = None, evidence_scale: GraphPosteriorEvidenceScale = 'latent-new', propagation_steps: int = 10, teleport_probability: float = 0.1, add_self_loops: bool = True, convolution_name: Literal['appnp', 'gcn'] = 'appnp') CUQGraphNeuralNetworkPredictor[In, Out][source]¶
Create a CUQ-GNN predictor with graph-refined node features.
Based on the committee-based graph uncertainty quantification approach of [DHullermeier24b], building on Graph Posterior Networks [SCG+21]. Graph convolutions refine node features before normalizing flows estimate class-conditional densities.
- Parameters:
input_encoder – Predictor applied to
data.xbefore graph refinement.latent_dim – Latent dimensionality used by the normalizing flow.
num_classes – Number of output classes.
encoder_dim – Output dimensionality of
input_encoder. Inferred when omitted.num_flows – Number of radial flow layers per class.
class_counts – Optional class-count prior. If omitted, counts are inferred from
data.train_mask.evidence_scale – Additive log-scale for feature evidence.
propagation_steps – Number of APPNP propagation steps.
teleport_probability – APPNP teleport probability.
add_self_loops – Whether APPNP should add self-loops.
convolution_name – Graph convolution used before density modeling. Supports
"appnp"and"gcn".
- Returns:
A CUQ-GNN predictor returning Dirichlet alphas.