torch_geometric.explain.metric.fidelity_curve_auc

fidelity_curve_auc(pos_fidelity: Tensor, neg_fidelity: Tensor, x: Tensor) → Tensor[source]

Returns the AUC for the fidelity curve as described in the “GraphFramEx: Towards Systematic Evaluation of Explainability Methods for Graph Neural Networks” paper.

More precisely, returns the AUC of

\[f(x) = \frac{\textrm{fid}_{+}}{1 - \textrm{fid}_{-}}\]
Parameters:

  • pos_fidelity (torch.Tensor) – The positive fidelity \(\textrm{fid}_{+}\).

  • neg_fidelity (torch.Tensor) – The negative fidelity \(\textrm{fid}_{-}\).

  • x (torch.Tensor) – Tensor containing the points on the \(x\)-axis. Needs to be sorted in ascending order.