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

Tab-Shapley: Identifying Top-k Tabular Data Quality Insights

AAAI 2025 · Learning theory and algorithms › Prediction and decision losses · approach: Shapley value attribution · 0 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.

Proposes Tab-Shapley, an unsupervised cooperative-game method that scores attributes and records by Shapley values with a closed form, then extracts top-k blocks of anomalous cells from tabular data.

Problem

Identifying prioritized top-k data quality insights (blocks of anomalous attributes and supporting records) in tabular data without labeled anomalies, while avoiding exponential Shapley computation and capturing attribute dependencies.

Approach

Defines a cooperative game over attributes whose characteristic function counts records in the union of their non-anomalous evidence sets; shows this super-additive game has a closed-form Shapley value (sum over records of 1/number of non-anomalous attributes), enabling O(mn) computation; then reorders the table and extracts top-k max-sum submatrices with Kadane's algorithm.

Key results

• Tab-Shapley captures more ground-truth anomalies than DIFFI in top-left 2x2, 4x4 and 6x6 blocks on most datasets, and is comparable to supervised SHAP (SHAP wins on several datasets at 4x4 and 6x6).
• On Forest Cover, Tab-Shapley outperforms both DIFFI and SHAP in the 6x6 to 12x12 submatrix anomaly counts in the reported table; on KDD Cup 1999, SHAP outperforms Tab-Shapley.
• Pearson correlations between Shapley values and the two criteria are high across datasets (e.g., 0.95 and 0.99 on KDD and Forest for Criteria 1 and 2 respectively, lowest 0.77 on Speech for Criteria 1).

Measured effects reported in the paper

Surprising findings

• Supervised SHAP outperforms the unsupervised Tab-Shapley on KDD Cup 1999 and on several datasets at 4x4 and 6x6 block sizes.
• Standard unsupervised Forest Cover results show SHAP capturing 0 anomalies in the 6x6 block while Tab-Shapley captures 4.

Limitations

• Relies on the quality of TabNet-based cell-level labels obtained by thresholding reconstruction errors with clustering.
• Evaluation on attribute-level ground truth only for a subset of datasets; KDD and Forest lack attribute-level ground truth.
• Human evaluation of the insights has not been conducted.

Experimental setup

Datasets: KDD Cup 1999 Forest Cover Arrhythmia Letter Ionosphere Speech Satimages WBC OPT SPECT Wine Vertebral

Benchmarks: KDD Cup 1999 Forest Cover Arrhythmia Letter Ionosphere Speech Satimages WBC OPT SPECT Wine Vertebral

Metrics: number of ground-truth PA anomalies captured in top-left k x Pearson correlation between Shapley values and evidence set

Baselines: DIFFI SHAP

Builds on: TabNet Kadane's algorithm DIFFI SHAP

Closest papers by meaning

• Beyond Individual Input for Deep Anomaly Detection on Tabular Data (ICML 2024 · similarity 0.86)
• Disentangling Tabular Data Towards Better One-Class Anomaly Detection (AAAI 2025 · similarity 0.85)
• RADAR: Benchmarking Language Models on Imperfect Tabular Data (NeurIPS 2025 · similarity 0.84)
• TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation (ICLR 2025 · similarity 0.84)
• Unsupervised Anomaly Detection for Tabular Data Using Deep Noise Evaluation (AAAI 2025 · similarity 0.84)
• Breaking the Quality–Privacy Tradeoff in Tabular Data Generation via In-Context Learning (NeurIPS 2026 · similarity 0.84)