Research map / Label-efficient and robust learning / Recognition and long-tail

LIMEFLDL: A Local Interpretable Model-Agnostic Explanations Approach for Label Distribution Learning

ICML 2025 · Label-efficient and robust learning › Recognition and long-tail · approach: Prototype and concept dictionary models

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.

Proposes LIMEFLDL, an extension of LIME that explains black-box label distribution learning models by solving a constrained ADMM-based linear surrogate over label distributions, with analytical solution, convergence/stability analysis, and LDL fidelity metrics.

Problem

Local interpretability of label distribution learning (LDL) models, which LIME (designed for single-label outputs) does not handle directly because of inter-label dependencies, computational cost of per-label sampling, and violation of label distribution constraints.

Approach

Fits one local linear surrogate with a feature attribution matrix over the whole label distribution (single local sampling), enforces sum-to-one and non-negativity constraints via ADMM with penalty terms, derives an analytical solution with convergence and stability bounds, and selects features with a KL-based scoring function.

Key results

• On Table 2 (RBF-LDL-LRR), LIMEFLDL gives lower Chebyshev/Clark/Canberra distances than PULIME on most tabular datasets, e.g., JAFFE Chebyshev 0.0338 vs 0.0461 and 3DFE 0.0348 vs 0.0583.
• Jaccard consistency is higher for LIMEFLDL on several datasets, e.g., JAFFE 0.7080 vs 0.5790 and Emotion6 0.3804 vs 0.1556 (Table 2).
• In human experiments, LIMEFLDL scored 3.1 vs PULIME 2.1 for accurate predictions and 3.1 vs 2.3 for incorrect predictions (1-5 scale).

Measured effects reported in the paper

Surprising findings

• PULIME's KL divergence can be negative because independently fitted per-label explanations may sum to more than 1, violating the distribution constraint.
• LIMEFLDL does not uniformly win: PULIME is better on several metrics for alpha, cold, and Scene, which the authors attribute to near-uniform label distributions.

Limitations

• Analytical solution relies on approximations and assumptions (e.g., Assumption 3.1 for the non-negativity penalty), so convergence and stability guarantees are partly heuristic.
• LIMEFLDL does not win on all datasets; PULIME performs better on several metrics for some datasets such as alpha, cold, and Scene.
• Human study is small (10 participants per part) and uses a self-rated 1-5 scale.

Experimental setup

Datasets: Yeast alpha Yeast cold JAFFE SBU 3DFE Emotion6 Natural Scene

Benchmarks: Yeast alpha Yeast cold JAFFE SBU 3DFE Emotion6 Natural Scene Image-JAFFE Image-Emotion6

Metrics: Chebyshev distance Clark distance Canberra distance KL divergence cosine similarity intersection similarity Jaccard index (top-K consistency)

Baselines: PULIME

Builds on: LIME ADMM PULIME (parallel LIME baseline)

Closest papers by meaning

• Adaptive-Grained Label Distribution Learning (AAAI 2025 · similarity 0.88)
• Approximately Correct Label Distribution Learning (ICML 2025 · similarity 0.87)
• Generative Calibration of Inaccurate Annotation for Label Distribution Learning (AAAI 2024 · similarity 0.87)
• Divide and Conquer: Learning Label Distribution with Subtasks (ICML 2025 · similarity 0.87)
• Predicting Label Distribution from Ternary Labels (NeurIPS 2024 · similarity 0.86)
• Concentration Distribution Learning from Label Distributions (ICML 2025 · similarity 0.85)