ATT accuracy

ATT accuracy asks whether the features assigned high importance by an explanation align with suitable reference evidence.

This is not the predictive accuracy of the black-box model. It is the accuracy of the attribution relative to a reference that identifies which tokens, regions, points, graph elements, or concepts should be relevant.

Required inputs

An ATT accuracy evaluation needs:

1. attribution scores for interpretable features;
2. a reference attribution or relevance annotation;
3. a rule for matching explanation features to reference features;
4. an aggregation metric and thresholding policy, where applicable.

The reference may come from human annotation, a semantic mask, domain-expert evidence, a controlled synthetic task, or another defensible source. The quality and uncertainty of that reference must be reported.

Modality-specific definitions

Modality	Attribution unit	Reference example	Suitable evaluation examples
Image classification	Pixels, superpixels, regions	Object or lesion mask	Coverage, weighted coverage, overlap, ranking against annotated regions
LLM / gSMILE	Tokens, words, phrases	Human-labelled influential text elements	ROC-AUC, precision, recall, F1, top-k agreement
Image editing	Instruction tokens or visual regions	Required edit terms or target edit region	Token-ranking agreement, edited-region overlap, expert judgement
Point cloud	Points, clusters, object parts	Part segmentation or annotated relevant cluster	Point/cluster overlap, coverage, top-k part agreement
KG-RAG	Entities, relations, paths	Gold evidence graph or supported reasoning path	Evidence precision/recall, ranking, path overlap
ConceptSMILE	Concepts and concept-relevant regions	Domain concept annotation	Concept-region alignment, concept detection agreement

LLM example from gSMILE

The gSMILE paper compares token attribution scores with ground-truth labels identifying relevant input words. It operationalises ATT accuracy with the ROC-AUC of attribution scores: a score of 1 indicates perfect ranking of relevant tokens above irrelevant tokens, while 0.5 corresponds to random ranking.

The gSMILE example compares token-level ground truth with generated token attributions. Source: Dehghani et al., Figure 6 and Section 4.4.1.

Image functions currently implemented in XWhy

XWhy currently provides image-specific spatial coverage functions through ImageCoverageMetrics:

from xwhy.metrics import ImageCoverageMetrics

coverage = ImageCoverageMetrics.calculate_coverage(
    explanation_image=explanation_map,
    semantic_mask=ground_truth_mask,
    class_of_interest=1,
)

weighted_coverage = ImageCoverageMetrics.calculate_weighted_coverage(
    explanation_image=explanation_map,
    semantic_mask=ground_truth_mask,
    class_of_interest=1,
)

coverage, weighted_coverage = ImageCoverageMetrics.evaluate_all(
    explanation_image=explanation_map,
    semantic_mask=ground_truth_mask,
    class_of_interest=1,
)

calculate_coverage() rewards selected explanation pixels within the target class and penalises selected pixels belonging to other labelled objects. calculate_weighted_coverage() additionally uses the magnitude of the explanation map.

The image-classification explainer can calculate these values when a ground-truth mask is supplied or when its configured segmentation pathway produces a semantic mask:

result = explainer.explain(
    instance="image.jpg",
    ground_truth_mask=mask,
)

print(result.coverage)
print(result.weighted_coverage)

Coverage is an image-specific ATT accuracy proxy

These functions do not implement a universal ATT accuracy metric. They assume spatial arrays with matching image dimensions and a semantic mask. They cannot be directly applied to LLM tokens, point-cloud clusters, knowledge-graph evidence, or concept lists.

Current support for other modalities

XWhy does not currently expose a public token-level AttAUC/F1 evaluator or a generic attribution-reference alignment class. For an LLM evaluation, the expected inputs would be an attribution score per token and a binary or graded reference relevance label aligned to the same tokenisation. That evaluator remains to be implemented in the package.

Similarly, point-cloud, image-editing, GraphRAG, and concept-level ATT accuracy require modality-specific feature alignment before a metric can be computed.

Reporting checklist

Report:

• who or what produced the reference attribution;
• annotation granularity and uncertainty;
• feature-alignment and tokenisation rules;
• whether negative attributions are included or transformed;
• threshold or top-k selection rules;
• class imbalance;
• both ranking and set/coverage results where possible.

Plausibility is not accuracy

An explanation may look reasonable without aligning with a defensible reference. Conversely, disagreement with one annotation does not automatically prove an explanation is wrong when the task admits multiple valid evidence sets.

Research basis

• gSMILE defines token-level ATT accuracy using annotated relevant text elements and AttAUC.
• ConceptSMILE evaluates whether concept attributions align with clinically meaningful retinal evidence.