class Explain(model: torch.nn.Module, adj_matrix: numpy.typing.NDArray, features: numpy.typing.NDArray)

Bases: object

An explainer class for instance-level explanations of Graph Neural Networks. Explanation is Probabilistic Graphical Model in a form of Bayesian network. This Bayesian network estimates the probability that node has the predicted role given a realization of other nodes.

How to build explanation:

explainer = Explain(model=model, A=A, X=X)
pgm_explanation = explainer.structure_learning(28)
Parameters:

  • model – (Module): The model to explain.

  • adj_matrix – (NDArray): Adjacency matrix of input data.

  • features – (NDArray): Feature matrix of input data.

pgm_conditional_prob(target: int, pgm_explanation: pgmpy.models.BayesianNetwork, evidence_list: List[str]) → float | None

Probability of target node, conditioned on the set of neighbours

Parameters:

  • target – (int) Index of node explained

  • pgm_explanation – (Bayesian Net) The Bayesian Net explaining the target

  • evidence_list – ([str]]) List of neighbours to condition on

Returns:

(float): The probability of the target node conditioned on ‘evidence_list’

structure_learning(target: int | None = None, top_node: int | None = None, num_samples: int = 20, pred_threshold: float = 0.1, child: bool | None = None) → pgmpy.models.BayesianNetwork

Learn structure of Bayesian Net, which represents pgm explanation of target node.

Parameters:

  • target – (str): Index of the node to be explained

  • top_node – (int, optional): The number of top the most probable nodes for Bayesian Net. If None, all nodes would be used (default: ‘None’)

  • num_samples – (int): The number of samples for data generation, which is used for structure learning, more number of samples – better learning.

  • pred_threshold – (float): Probability that the features in each node is perturbed (default: 0.1)

  • child – (bool, Optional): If False or None, no-child constraint is applied (default: None)

Returns:

(BayesianNetwork): Pgm explanation in Bayesian Net form