Research map / Learning theory and algorithms / Prediction and decision losses

Accelerating the Global Aggregation of Local Explanations

AAAI 2024 · Learning theory and algorithms › Prediction and decision losses · approach: Provable post-hoc explanations · 2 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Accelerates global aggregation of Anchor local explanations into top-k impactful words via an anytime algorithm and lossy Anchor hyperparameter changes, and proposes a probabilistic aggregation (G_pr) that discounts noisy frequent words.

Problem

Global aggregation of local Anchor token explanations over a dataset is computationally prohibitive (days for 10k documents), and common aggregations (sum, average) are biased toward frequent noisy or rare words.

Approach

A probabilistic generative model with anchor and non-anchor word distributions, estimated by MLE, yields a per-word aggregation score q*(w,c); combined with an anytime top-k algorithm using pseudo-scores, candidate filtering, and reduced Anchor sampling (fewer masking suggestions, lower precision threshold, higher confidence delta).

Key results

• Optimizations accelerate global aggregation of Anchor by up to 30x, reducing computation from hours to minutes while incurring a very mild loss of quality.
• The probabilistic aggregation G_pr consistently begins with the steepest AOPC curve and exceeds baselines; G_av is inferior to the other aggregations.
• Counterfactual sentences with top G_pr terms reduce classifier accuracy by roughly 50% when appended to opposite-label documents.

Measured effects reported in the paper

Surprising findings

• Increasing the confidence parameter delta improves both running time and quality of G_pr, suggesting aggregation compensates for reduced per-instance Anchor confidence.
• Reducing the MLM suggestion count from 500 to 50 in masking generally did not harm and sometimes improved quality.

Limitations

• Focuses only on single-word terms; multi-word terms left for future work.
• Combining all optimizations lowers AOPC quality in most experiments, with the confidence optimization causing the largest quality loss.
• Experiments restricted to short documents (up to 200 characters) and few classification tasks.

Experimental setup

Datasets: Coronavirus tweets The Social Dilemma tweets Amazon reviews: Toys & Games (spam) Amazon reviews: Home & Kitchen (spam)

Benchmarks: Coronavirus tweets The Social Dilemma tweets Amazon reviews: Toys & Games (spam) Amazon reviews: Home & Kitchen (spam)

Metrics: AOPC^k (adapted to top-k terms) running time ratio of shared terms accuracy drop under counterfactual sentence addition

Baselines: sqrt aggregation (G_sq) average importance (G_av) G_h entropy aggregation class-probability baseline (G_b) inverse G_pr (G_rev) G_av with rare-word filtering

Builds on: Anchor AOPC LIME SHAP

Closest papers by meaning

• GnnXemplar: Exemplars to Explanations - Natural Language Rules for Global GNN Interpretability (NeurIPS 2025 · similarity 0.83)
• Local vs. Global Interpretability: A Computational Complexity Perspective (ICML 2024 · similarity 0.83)
• Short-Context Dominance: How Much Local Context Natural Language Actually Needs? (NeurIPS 2026 · similarity 0.82)
• Anchored Diffusion Language Model (NeurIPS 2025 · similarity 0.82)
• Test-Time Training on Nearest Neighbors for Large Language Models (ICLR 2024 · similarity 0.82)
• Golden Layers and Where to Find Them: Improved Knowledge Editing for Large Language Models Via Layer Gradient Analysis (NeurIPS 2026 · similarity 0.82)