Pareto Interpretation Eval

pareto-interpretation-eval

Synthesizing Pareto-Optimal Interpretations for Black-Box Models — Torfah et al. (2021) (arXiv:2108.07307, 2021)

What this evaluates

This evaluation probes a model's ability to synthesize interpretable surrogate models (decision diagrams) that explicitly trade off prediction fidelity against structural simplicity. It measures how well a method can navigate the Pareto front between accuracy and explainability without collapsing them into a single weighted objective.

Datasets

• Airplane Perception Module (AP) — total ?; splits: test (-1)
• Bank Loan Predictor (BL) — total ?; splits: test (-1)
• Theorem Prover Solvability Predictor (TP) — total ?; splits: test (-1)

Metrics

• accuracy (primary) — range: [0, 1]
  • Standard classification accuracy: fraction of correctly predicted labels out of total samples in the evaluation set.
• explainability_score — range: other
  • A structural measure favoring decision diagrams with fewer nodes (size) and predicates with fewer branchings. Lower raw values indicate higher explainability.

Input / output format

Input: Feature vectors specific to each benchmark: (1) time of day, cloud types, and initial positioning for AP; (2) age and income for BL; (3) percentage of unit clauses and average clause length for TP.

Output: A decision diagram (interpretation) representing a synthesized trade-off solution between correctness and explainability.

Scoring recipe

def evaluate(predictions, gold):
    accuracy = sum(p == g for p, g in zip(predictions, gold)) / len(gold)
    dd = predictions.diagram
    dd_size = count_nodes(dd)
    branching = count_branchings(dd)
    explainability = -(dd_size + branching)  # Higher is better
    return {'accuracy': accuracy, 'explainability': explainability}

Common pitfalls

• Comparing with single-objective methods (e.g., MinDS) is unfair because they do not guarantee exploration of the full Pareto-optimal space.
• Sample sizes are not fixed; they are derived from statistical confidence (δ) and error margin (ε), so direct comparison across different settings requires normalization.
• Explainability is defined structurally (decision diagram size and branching) rather than via standard human-study metrics, limiting direct cross-domain benchmarking.

Evidence (verbatim from paper)

We used accuracy for correctness, and explainability measure that favored decision diagrams of smaller size and predicates with a fewer number of branchings.

Citation

@misc{torfah2021pareto,
  title={Synthesizing Pareto-Optimal Interpretations for Black-Box Models},
  author={Torfah et al. (2021)},
  year={2021},
  note={arXiv:2108.07307}
}

• arXiv: 2108.07307