Shapley Integrated Gradients: Unified Attributions

• Shapley Integrated Gradients (SIG) is a unified feature attribution framework that combines the fairness of Shapley values with the efficiency of Integrated Gradients.
• SIG methods overcome baseline selection bias and the limitations of a single interpolation path by averaging over multiple baselines and coalitions.
• Variants like baseline sampling, IGCS, Compensated IG, and Group SIG illustrate SIG's versatility with reduced error and computational cost.

Shapley Integrated Gradients (SIG) unify the axiomatic rigor of Shapley value attributions with the computational efficiency of Integrated Gradients (IG), establishing a family of techniques for feature contribution analysis in complex models. SIG methods address the limitations of standard IG—most notably, the arbitrariness and potential bias of baseline selection, and the failure of a single interpolating path to replicate the combinatorial averaging inherent to the discrete Shapley value. Recent developments cover diverse theoretical frameworks and practical implementations, including proportional coalition sampling, “soft” cohort averaging, compensation via Shapley sampling, and group-based extensions. This article surveys the conceptual foundations, representative algorithms, theoretical properties, comparative validation, and key variants of Shapley Integrated Gradients.

1. Theoretical Foundations and Motivation

Shapley Integrated Gradients emerge from the intersection of two paradigms for feature attribution: the Shapley value, which uniquely satisfies fairness axioms for contributions in cooperative settings, and Integrated Gradients, which approximate feature attributions by integrating local model gradients along a path from a reference baseline to an input. IG naturally solves the so-called “Aumann–Shapley” cost-sharing problem via a straight-line interpolation in $\mathbb{R}^n$, yielding path-integral attributions that sum to the model’s output difference between input and baseline.

However, IG’s reliance on a single baseline and a unique path omits the combinatorial structure of classical Shapley value, which considers all $2^n$ feature coalitions or, equivalently, all $n!$ permutations. This leads to attributions that may be biased or unintuitive, especially if the baseline does not correspond to a meaningful absence of information or fails to be “near-zero” in the relevant domain. In contrast, the Shapley value for an input $x$ and model $F$ is

$$\phi_i(v) = \sum_{S \subseteq N \setminus \{i\}} \frac{|S|!(|N|-|S|-1)!}{|N|!} \left[ v(S \cup \{i\}) - v(S) \right],$$

where $v(S)$ evaluates $F$ with features in $S$ set to $x$ and others to baseline. The computational cost of evaluating all coalitions is prohibitive for $n \gtrsim 20$. SIG variants advance efficient approximations that inherit much of the Shapley value’s interpretability and fairness guarantees (Liu et al., 2023, Hama et al., 2022, Shanbhag et al., 2021, Tachikawa et al., 2018).

2. Variants and Algorithmic Frameworks

Several forms of Shapley Integrated Gradients have been developed, each suited to a distinct context:

Variant	Key Idea	Reference
SIG via Baseline Sampling	Averaging IG from coalitional baselines in Shapley proportions	(Liu et al., 2023)
Cohort SIG (“IGCS”)	Path integration over soft cohort averages	(Hama et al., 2022)
Compensated IG (C-IG)	Baseline correction using Shapley sampling	(Tachikawa et al., 2018)
Group SIG	Path-integral Shapley for grouped features/distributions	(Shanbhag et al., 2021)

SIG via Baseline Sampling: SIG (Liu et al., 2023) interprets each coalition (subset $S \subseteq N$) as a distinct baseline $d_S$ defined by $d_{S,i}=x_i$ for $i \in S$, $d_{S,i}=x'_i$ otherwise. Proportional sampling draws baselines according to the weights $w(k)$ corresponding to the coalition size, then averages the IG attributions computed from these sampled baselines. The resulting estimate is unbiased and converges to the Shapley value with $O(1/B)$ variance as the number of baselines $B\to\infty$, with computational cost $O(B n^2)$ (much less than exact Shapley’s $O(2^n)$ scaling).

Cohort SIG (“IGCS”): IGCS (Hama et al., 2022) softens the discrete cohort Shapley approach by introducing a weight vector $z \in [0,1]^d$ and computing IG along the main diagonal in this space. The “cohort mean” $\nu(z)$ acts as a differentiable version of the Shapley value computation counting over similar instances. For high $d$, IGCS produces attributions nearly indistinguishable from cohort Shapley at linear cost $O(nd)$, and unlike standard IG, it is compatible with binary features.

Compensated IG: SIG or C-IG (Tachikawa et al., 2018) augments IG attributions—calculated from any (possibly suboptimal) baseline—with a feature-wise compensation vector computed as the difference between Shapley sampling and IG on a held-out calibration set. This scheme maintains efficiency (cost intermediate between IG and Shapley sampling) while correcting baseline bias and preserving the required axioms.

Group SIG: GroupSIG (Shanbhag et al., 2021) generalizes standard IG to group-shaped attribution, combining the path-integral idea with functionals $D$ representing distributional or aggregate differences (e.g., mean shift, Wasserstein-$1$ distance) over cohorts or feature groups. This framework supports attribution of prediction drift between datasets and localizes changes to groups of features or data points.

3. Axiomatic Properties and Theoretical Guarantees

SIG variants leverage the path-integral structure of IG to inherit or approximate the classic Shapley axioms:

• Efficiency: The sum of attributions equals the total function difference,

$$\sum_i SIG_i = F(x) - F(x') \quad \text{(or } D(G(F(S_1)),G(F(S_2)))\text{ for group variants)}.$$

• Dummy: Attributions for features with no effect on the output are zero.
• Linearity: Attributions for a linear combination of models are the corresponding combination of their attributions.
• Symmetry: Identically contributing features receive equal attributions, often subject to model architecture (e.g., convolutional symmetry in C-IG).
• Proportionality and Group Additivity: Group SIG preserves additive structure for grouped features, supporting coarse-to-fine interpretability.

Unbiasedness (for stochastic sampling approaches as in (Liu et al., 2023)) and convergence to the true Shapley value as $B \to \infty$ are formally established. For cohort-based IGCS (Hama et al., 2022), multilinearity arguments show that over most of the domain, the attributions are close to cohort Shapley values with error $O(n^2 \epsilon^2)$, exponentially small in the number of features.

4. Algorithmic Complexity and Practical Considerations

SIG methods balance fairness/interpretable axioms against tractability:

• SIG via Sampling: Cost is $O(B n^2)$, affordable for moderate dimensions ($n \lesssim 1000$) if batched on GPU (Liu et al., 2023).
• IGCS: Cost is $O(ndR)$, where $R$ is the number of quadrature/integration points ($R\approx 50-200$ suffices) (Hama et al., 2022).
• Compensated IG: Initial calibration requires $K M$ IG and $K P n$ Shapley sampling evaluations, but per-input cost is $O(Mn)$ for IG alone (Tachikawa et al., 2018).
• GroupSIG: Complexity is $O(K)$ backpropagations for $K$ steps along the path (typically $50-200$) (Shanbhag et al., 2021).

For all SIG methods, baseline and group (“player”) design directly influence attribution granularity and runtime. For images, features are often grouped as patches; for text or categorical data, appropriate feature representations must be defined.

5. Empirical Validation and Comparative Evaluation

Empirical studies validate SIG across domains and compare with standard IG and sampling-based Shapley estimators:

• GridWorld RL: SIG achieves high Spearman’s $\rho$ ($\approx 0.9–0.95$) when correlating its rank order with ground-truth Shapley values, outperforming single-baseline IG and mean/random baselines with lower variance (Liu et al., 2023).
• Image Classification: On image tasks, SIG produces sharper and more semantically meaningful saliency maps (matching human intuition in facial expression and ImageNet tasks) compared to standard IG. Quantitatively, iAccuracy curves for SIG closely match those of classic IG but provide improved localization (Liu et al., 2023).
• High-dimensional Variable Importance: For 16D physics and 1024D chemistry models, IGCS nearly matches cohort Shapley on ABC metrics, with performance surpassing Monte Carlo Shapley at equivalent cost (Hama et al., 2022).
• EEG Classification: Compensated IG returns attributions with Spearman rank correlations $0.97–0.99$ against Shapley sampling (pseudo-ground-truth), much higher than standard IG (frequently $<0.70$), and more reliable spatial maps (Tachikawa et al., 2018).
• Prediction Drift: GroupSIG robustly localizes drifts in high-dimensional tabular data, outperforming tree-based or non-path approaches in simulated pipeline bug scenarios (Shanbhag et al., 2021).

These results demonstrate practical benefits of SIG, particularly in robustness to baseline choice and alignment with human domain knowledge.

6. Application Scenarios and Limitations

SIG is applicable across supervised learning domains: image, time series, text, and structured data, provided a suitable differentiable model or functional is available. It is compatible with arbitrary data modalities through appropriate “player” design—grouping of pixels, patching features, or defining binary or categorical similarity.

Limitations include:

• Absent features for very small numbers of samples ($B$) yield noisy attributions in sampling-based SIG.
• Patch-based or grouped features risk smoothing out fine detail.
• GroupSIG requires differentiable $D$, $G$, and $F$; non-differentiable metrics (e.g., KS-statistic) are incompatible without smoothing (Shanbhag et al., 2021).
• For “on-manifold” explanations, care must be taken in similarity and group specification to ensure meaningful attribution.
• Standard IGCS and other data-driven forms depend on a sufficient number of observations; small $n$ or high dimensionality can stress approximation quality.

7. Relationship to Existing Attribution Methods

SIG can be seen as a unification and generalization of existing approaches:

• Integrated Gradients: Special case when a single baseline and output-difference are used; fails to capture combinatorial attribution.
• Shapley Value (sampling): Statistical SIG methods approximate the Shapley value at dramatically lower computational cost.
• Cohort Shapley: IGCS recovers CS closely in high dimension, handling both continuous and binary predictors (Hama et al., 2022).
• GroupSHAP and KernelSHAP: SIG offers a gradient-based and more efficient alternative when models are differentiable and grouping is meaningful (Shanbhag et al., 2021).

SIG thus synthesizes the strengths of axiomatic, combinatorial, and gradient-based variable importance, forming a versatile core in modern explanatory machine learning.