Pgexplainer Gnn Eval

pgexplainer-gnn-eval

Parameterized Explainer for Graph Neural Network — Luo et al. (2020) (arXiv:2011.04573, 2020)

What this evaluates

Evaluates the predictive accuracy of GNN models and the effectiveness of post-hoc explanation methods on synthetic and real-world graph classification and node classification tasks. It probes whether a parameterized explainer can learn global explanatory motifs end-to-end and generalize inductively without retraining.

Datasets

• BA-Shapes — total ?; splits: train (-1), val (-1), test (-1)
• BA-Community — total ?; splits: train (-1), val (-1), test (-1)
• Tree-Cycles — total ?; splits: train (-1), val (-1), test (-1)
• Tree-Grid — total ?; splits: train (-1), val (-1), test (-1)
• BA-2motifs — total ?; splits: train (-1), val (-1), test (-1)
• MUTAG — total ?; splits: train (-1), val (-1), test (-1)

Metrics

• Accuracy (primary) — range: percent
  • Standard classification accuracy: the proportion of correctly predicted labels out of the total number of instances.

Input / output format

Input: Graph-structured data containing node features, adjacency matrices, and node or graph-level labels for classification tasks.

Output: Predicted class labels for nodes or graphs, and binary edge/node importance scores for explanation methods.

Scoring recipe

def compute_accuracy(predictions, gold_labels):
    correct = sum(1 for p, g in zip(predictions, gold_labels) if p == g)
    return correct / len(gold_labels) * 100

Common pitfalls

• GNN models must be fully trained first and then shared with all post-hoc explanation methods rather than training them jointly.
• The dataset split is fixed at 80/10/10 for train/validation/test across all datasets.
• PGExplainer requires careful temperature annealing (tau0=5.0 to tauT=2.0) to balance discrete sampling and backpropagation optimization.

Evidence (verbatim from paper)

We follow GNNExplainer to split train/validation/test with 80/10/10% for all datasets. Each model is trained for 1000 epochs. The accuracy performances of GNN models are shown in Table 3.

Citation

@misc{luo2020pgexplainer,
  title={Parameterized Explainer for Graph Neural Network},
  author={Luo et al. (2020)},
  year={2020},
  note={arXiv:2011.04573}
}

• arXiv: 2011.04573