TabSHAP

1. Problem Statement

The paper addresses the critical interpretability gap in applying Large Language Models (LLMs) to structured tabular data. While LLMs fine-tuned on serialized tabular data (e.g., converting rows to natural language prompts) have shown superior performance on heterogeneous datasets compared to traditional tree-based models (like XGBoost), they lack faithful interpretability required for high-stakes domains (healthcare, finance).

Existing interpretability methods fail in this context due to three main issues:

1. Inappropriate Granularity: Standard token-level attribution (deleting sub-word tokens) corrupts the semantic integrity of tabular fields (e.g., deleting "5" from "Age: 50" creates invalid syntax or hallucinations).
2. Oversimplified Metrics: Most methods rely on global linear proxies (e.g., fitting logistic regression) or scalar probability shifts (prediction flips), which fail to capture the model's full probabilistic uncertainty and non-linear decision logic.
3. Generative Nature: LLMs distribute probability mass over vast vocabularies, making it difficult to extract stable class probabilities for attribution without specific aggregation strategies.

2. Methodology: TABSHAP

The authors propose TABSHAP, a model-agnostic, post-hoc interpretability framework designed specifically for LLM-based tabular classifiers. The pipeline consists of four core components:

A. Tabular Serialization & Atomic Masking

• Serialization: Tabular rows are converted into concise, space-delimited key:value strings (e.g., age:50 workclass:private).
• Atomic Masking: Instead of deleting individual tokens, TABSHAP treats each key:value pair as an atomic unit. When a feature is "removed" for a coalition, the entire string is omitted from the prompt. This preserves prompt syntax and prevents model hallucinations, ensuring the perturbation corresponds to a meaningful feature absence.

B. Verbalizer-Style Class Aggregation

• To handle the generative nature of LLMs, where a class label (e.g., "Yes") might be represented by multiple tokens ("Yes", " yes", "ye"+"s"), TABSHAP employs a Verbalizer Mapping.
• It aggregates probability mass from the top-K most probable next tokens that match the target class set.
• The aggregated probabilities are then normalized to form a valid class distribution $P(c|x)$, ensuring stability even for distilled or quantized models.

C. Distributional Similarity via Jensen-Shannon Divergence (JSD)

• Rather than measuring if a prediction label flips, TABSHAP measures the distributional impact of a feature.
• It computes the Jensen-Shannon Divergence (JSD) between the full-input class distribution ($P_{full}$) and the distribution under a feature coalition ($P_S$).
• The value function $v(S)$ is defined as a normalized similarity score:
  $$v(S) = 1 - \min\left(\frac{JSD_{nat}(P_{full} \| P_S)}{\ln 2}, 1\right)$$
• This captures shifts in model confidence across all classes, not just the top prediction.

D. Sampled-Coalition Estimation

• Since exact Shapley value calculation is intractable for LLMs, TABSHAP uses a Monte Carlo sampled-coalition estimator (adapted from TokenSHAP).
• It evaluates a subset of feature coalitions (including all "leave-one-out" sets and random subsets) to approximate feature importance ($\phi_j$).
• The final score is the difference in average similarity between coalitions containing the feature and those excluding it.

3. Key Contributions

1. Distributional Attribution: Introduces JSD as the primary metric for feature importance, capturing the full shift in model confidence rather than simple label flips.
2. Feature-Level Atomic Masking: Solves the semantic corruption problem by masking entire key:value fields instead of sub-word tokens, ensuring perturbations are semantically valid.
3. Verbalizer Aggregation: Provides a robust method to aggregate multi-token class probabilities, stabilizing attributions for generative models.
4. Empirical Validation: Demonstrates that LLMs and tree-based models rely on fundamentally different decision logic, with LLMs exhibiting strong biases toward semantic/text-rich features.

4. Experimental Results

The framework was validated on two benchmarks: Adult Income (binary classification, 14 features) and Heart Disease (multi-class, 13 features), using a DeepSeek-R1-Distill-Llama-8B model fine-tuned via QLoRA.

• Faithfulness (Deletion Curves):
  • TABSHAP significantly outperformed random baselines and XGBoost proxies.
  • When features ranked by TABSHAP were sequentially removed, the model's confidence in the original prediction dropped precipitously, indicating that TABSHAP correctly identifies the features driving the decision.
• Metric Ablation:
  • The authors compared JSD against KL Divergence and L1 distance. JSD-based rankings generally provided the most faithful attribution curves, confirming the importance of distributional similarity over simple distance metrics.
• LLM vs. XGBoost Comparison:
  • Rank Correlation: Moderate Spearman correlation ($\rho \approx 0.65$) between TABSHAP and TreeSHAP rankings.
  • Semantic Bias: While XGBoost prioritized precise numerical splits (e.g., "Hours-per-week"), TABSHAP revealed that the LLM assigns higher importance to semantic, text-rich features (e.g., "Occupation", "Education"). This highlights that LLMs leverage semantic priors in ways linear proxies cannot detect.

5. Significance and Limitations

Significance:
TABSHAP bridges the gap between the powerful generalization of LLMs and the strict interpretability requirements of critical decision-making systems. By providing faithful, instance-level attributions, it enables the responsible deployment of LLMs in healthcare and finance, moving beyond "black box" predictions to explainable AI.

Limitations:

• Computational Cost: Requires multiple forward passes per instance, making it expensive in wall-clock time (though sampling caps mitigate this).
• Approximation: Uses sampled coalitions rather than the full exponential Shapley formula; faithfulness depends on sampling parameters ($r$, $C_{max}$).
• Scalability: Currently validated on low-dimensional datasets ($M \approx 14$); high-dimensional tabular settings remain unexplored.
• Interactions: Reports marginal attributions and does not explicitly model higher-order feature interactions.
• Architecture Specificity: Validated on a single backbone; behavior may vary across different LLM architectures.