TL;DR
SIM-Shapley introduces a stable, efficient method for approximating Shapley values in high-dimensional models, significantly reducing computation time while maintaining attribution accuracy, with broad applicability to sample average problems.
Contribution
It presents a novel stochastic iterative approach for Shapley value approximation that improves stability and efficiency over existing methods.
Findings
Reduces computation time by up to 85% compared to baselines.
Maintains comparable feature attribution quality.
Proven linear Q-convergence and theoretical variance analysis.
Abstract
Explainable artificial intelligence (XAI) is essential for trustworthy machine learning (ML), particularly in high-stakes domains such as healthcare and finance. Shapley value (SV) methods provide a principled framework for feature attribution in complex models but incur high computational costs, limiting their scalability in high-dimensional settings. We propose Stochastic Iterative Momentum for Shapley Value Approximation (SIM-Shapley), a stable and efficient SV approximation method inspired by stochastic optimization. We analyze variance theoretically, prove linear -convergence, and demonstrate improved empirical stability and low bias in practice on real-world datasets. In our numerical experiments, SIM-Shapley reduces computation time by up to 85% relative to state-of-the-art baselines while maintaining comparable feature attribution quality. Beyond feature attribution, our…
Peer Reviews
Decision·Submitted to ICLR 2026
1. The paper is well-written, with a clear structure that allows readers to easily follow the authors' reasoning. 2. The comparison between SIM-Shapley and other Shapley value computation methods is thorough and clear, enabling readers to readily grasp the paper's contributions. 3. Relevant experiments effectively demonstrate the superiority of the proposed method, particularly in accelerating Shapley value computation. 4. The paper addresses interpretability, a critical issue in deep learning m
The paper is relatively comprehensive, and there are no major fundamental issues. However, two minor points require attention: 1. **Paper Formatting:** The authors are requested to review the paper's formatting. For instance, on Page 21 of the supplementary materials, some figures/icons clearly exceed the paper's margins. 2. **Evaluation Metrics:** The paper primarily uses the error between estimated Shapley values and the ground truth to evaluate performance. Readers are curious about how the p
1. The proposed method covers both local and global explanations. 2. Clear reformulation of KernelSHAP as a constrained stochastic optimization with a simple EMA update and a closed-form per-iteration solution. 3. The method stays model-agnostic.
1. The authors claim that their method fundamentally re-conceptualizes SV computation. However, the core of their process is a series of mature stochastic optimization techniques. 2. The early stopping method (Eq.11) is heuristic and has no sensitivity analysis to the parameter epsilon. 3. The choice of the parameter xi in Eq. 12 is crude and lacks analysis. 4. The authors claim that the method can be used for Shapley Interactions, but provide no experiments to demonstrate this.
N/A
Big: 1. I think the biggest structural issue in this paper is how they discuss and compare to prior work. Across the discussions and experiments, they consider different subsets of the estimators FastSHAP, SimSHAP, LeverageSHAP, and KernelSHAP. The first issue is that they selectively compare to these algorithms e.g., they'll sometimes ignore LeverageSHAP or FastSHAP depending on the experiment. The bigger issue is that they don't consider other SOTA estimators which include PermutationSHAP, an
Code & Models
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
