acceptodds
Under review as a conference paper at ICLR 2027

Feature Attribution from Multi-Source Data with Heterogeneous Observation Patterns

Abstract

In multi-source learning, different sites typically observe different feature subsets, so pooling data across sites produces blockwise missingness and distributional shift. Shapley-based feature attribution has become a standard tool for model interpretability, yet existing estimators require fully observed evaluation data. We propose DASH (Data-fused Attribution via Shapley), which derives the influence function of the constrained weighted-least-squares Shapley estimator and uses it to construct site-specific control variates from partially observed auxiliary sites, reducing variance without imputing missing features. A permutation-based screening step excludes incompatible sources, and data-adaptive calibration optimally weights each site. In simulations, DASH achieves – lower MSE than the single-site estimator and – lower error than imputation baselines, with screening power at moderate misalignment. On environmental monitoring and multi-center clinical data, DASH reduces error by – over the single-site baseline, while imputation can increase error in the clinical setting.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Multi-source data fusion, blockwise missing data, influence functions, control variates, variance reduction, Shapley values, feature attribution, model interpretability

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_mN5ROpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes DASH for estimating Shapley feature attributions from multi-source data with heterogeneous feature coverage. The method derives the influence function of a constrained WLS Shapley estimator and uses it to construct control variates from partially observed auxiliary sites. It also introduces permutation-based source screening and variance-based calibration weights. Simulations and two real-data experiments show improvements over single-site and imputation-based estimators. I would be willing to reconsider my rating if the authors satisfactorily address these concerns.

Critical strengths
  1. The paper addresses a relevant problem in multi-source model interpretation, where different sites observe different subsets of features.

  2. The influence-function-based construction provides a reasonable way to incorporate partially observed auxiliary data without imputing missing features.

  3. The experimental results are encouraging, particularly the improvements over single-site and imputation-based estimators.

Critical weaknesses
  1. The method relies on several estimated quantities, including density ratios, influence functions, and control variates, but the theoretical treatment of their combined estimation errors is limited.

  2. The permutation-based screening procedure is an important component of DASH, yet its validity under unequal source distributions and estimated importance weights is not established.

  3. The numerical experiments mainly consider low-dimensional settings with moderate feature missingness. It is unclear how stable the proposed method remains when the number of features or missing-data patterns increases.

Questions for authors

1. In Section 3.3 (Page 5), the permutation p-value is defined using the indicator I(Tr≥Tr(b))I(T_r\geq T_r^{(b)}). Since larger values of TrT_r indicate greater disagreement between sites, should the inequality be reversed? Please clarify whether this is a typo or also affects the implementation.

2. In Algorithm 1 (Appendix B, Step 11), the influence function is computed as Σ−1[gw+λw1]\Sigma^{-1}[g_w+\lambda_w\mathbf{1}], whereas Eq. (10) gives Σ−1gw+λw1\Sigma^{-1}g_w+\lambda_w\mathbf{1}. These expressions are generally different. Which expression is used in the experiments?

3. In Appendix A, the influence-function expansion under covariate shift assumes ∥w^−w∥∞=op(n−1/2)\|\widehat w-w\|_\infty=o_p(n^{-1/2}). This seems demanding for density ratios estimated using a gradient-boosted classifier. Is this rate plausible under the proposed implementation, and how sensitive are the results to density-ratio estimation errors?

4. The calibration weights in Eqs. (17)–(18) are derived under no covariate shift, but the same weights are retained when importance weighting is introduced in Eq. (16). Are these weights still variance-optimal under covariate shift? If not, how much efficiency could be lost?

5. In the NACC application, the missing features are artificially removed from an initially complete dataset. Could the authors clarify how the method would perform with naturally occurring blockwise missingness, where missingness patterns may be associated with site characteristics or patient populations?

Reviewer_ksZSOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposed DASH, a method for estimating global Shapley feature attributions from multi-site data with blockwise missingness. By deriving the influence function of the constrained WLS Shapley estimator and constructing site-specific control variates, DASH reduces variance without imputation, avoiding the systematic coalition-dependent bias that imputation introduces.

Critical strengths
  1. The topic is valuable. Estimating global feature attributions when different data sources observe different feature subsets is practically relevant. Predicting the Shapley estimator’s influence contributions from each site’s observed variables provides a natural way to use auxiliary data without reconstructing every missing feature. The resulting construction is interpretable and could be useful when complete labeled data are scarce.

  2. The simulation experiments provide encouraging evidence for the approach. Table 1 shows improvements over the LC+IPW baseline, so the reported gains are not entirely explained by correcting covariate shift. The improvements across several prediction models in the two applications also support the practical value of the fusion strategy.

Critical weaknesses
  1. The theoretical justification does not fully cover the implemented estimator under covariate shift. Equation (17) calibrates using unweighted control covariances, whereas Equation (16) averages density-weighted controls. Variance-optimal calibration generally requires covariances of these weighted quantities. Appendix D establishes optimality without shift, so retaining the same rule does not establish variance reduction in the heterogeneous setting central to the paper. Separately, Appendix A assumes ∥w^−w∥∞=op(n−1/2)\|\widehat w-w\|_\infty=o_p(n^{-1/2}), without showing that the proposed classifier-based density-ratio estimator satisfies this strong condition. These gaps leave uncertainty about whether the theoretical efficiency argument applies to the procedure actually evaluated.

  2. The experimental settings are relatively simple. The main simulations use only 10 features with an independent Gaussian covariate structure, while the two real-data applications contain only 9–10 features and use 300 complete labeled observations. The runtime study also extends only to 12 features. Although the paper includes sample-size ablations and uses sizable auxiliary datasets, these experiments provide limited evidence about performance when the complete labeled sample size and feature dimensionality increase together, particularly when the number of features becomes comparable to or exceeds the number of complete labeled observations. This matters because the method relies on estimating conditional influence functions and calibration coefficients, whose accuracy and stability may deteriorate as dimensionality increases.

Questions for authors

1.What is the main methodological step beyond applying DEFUSE-style control-variate fusion to the Shapley influence vector? Please distinguish the components inherited from existing fusion methods from those requiring new arguments or providing a specific computational advantage.

2.Could the authors extend both the synthetic and real-data evaluations to larger sample sizes and substantially higher feature dimensions, including high-dimensional setting?

Reviewer_DK9DOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

DASH estimates global Shapley feature attributions for a fixed predictor using a small labeled-complete sample, labeled auxiliary sites with blockwise-missing features, and a larger unlabeled-complete target sample. The paper expresses the SAGE-style attribution vector as a constrained weighted least-squares solution, derives its observation-level influence function, and regresses this influence vector on each auxiliary site's observed variables to construct control variates. Importance weighting addresses covariate shift, a permutation procedure screens sources, and calibration combines their corrections. Simulations and controlled missingness experiments on Beijing air-quality and NACC data report lower attribution MSE than single-source and imputation approaches.

Critical strengths
  1. The method tackles a practically relevant attribution problem without requiring full reconstruction of missing feature blocks. The influence-vector construction connects the constrained Shapley problem to control-variate fusion in a transparent way, and cross-fitted regression provides a useful implementation for learning the auxiliary information.

  2. The evaluation reports bias–variance decompositions rather than only aggregate MSE, includes an importance-weighted single-site baseline, and tests misaligned-source screening. The Beijing results remain favorable even against the weighted baseline, while the NACC experiment demonstrates benefits in the simpler aligned setting. These results support further development of the approach.

Critical weaknesses
  1. The practical covariate-shift estimator is not covered by the main influence-function and optimal-calibration argument. Appendix A assumes density-ratio error negligible at first order, giving ||w_hat−w||_infinity=o_p(n^−1/2) as a sufficient condition, but no justification is supplied for the fitted classifier. Appendix D derives calibration without shift and explicitly retains that rule under shift; minimizing the variance of Eq. (16) instead requires covariances of the weighted correction terms. Assumption 1 also states LM–UC alignment without explicitly stating the LC–UC conditional-outcome alignment needed by Eq. (8). The acknowledged lack of exchangeability in screening adds another limitation to the claimed protection against incompatible sites. These issues matter because heterogeneous sources are the central setting.

  2. The empirical evidence does not yet cleanly establish the incremental gain or the target of evaluation. In Table 1, the MSE gains over the relevant importance-weighted single-site baseline are approximately 1.77×, 1.68×, and 1.25×, appreciably smaller than the headline comparison against the unweighted baseline. In Appendix E, predictors are trained on all non-LC observations, while the reference is computed from “held-out training observations” for Beijing and all non-LC patients for NACC. The overlap between model training, auxiliary estimation, and reference evaluation is unclear, and the Beijing reference is not explicitly restricted to the UC target population. This leaves uncertainty about the fixed-model target and independence assumed in the analysis.

Questions for authors
  1. What are the complete LC–UC identification and overlap assumptions? For the implemented classifier-based weights, can you derive the applicable first-order expansion and weighted calibration, or explicitly limit the theoretical claims to known weights/no shift and provide an oracle-weight versus estimated-weight ablation? Please also clarify whether Algorithm 1 regresses the weighted or unweighted influence vector before applying Eq. (16).

  2. Please specify the row-level membership and overlap of model-training, LC/LM estimation, UC background, and reference-evaluation sets. Can you report results using an independent labeled sample from the UC population as reference, including improvements over Single+IPW and screening calibration under stronger aligned covariate shift? These checks would materially affect my recommendation.

1 comment

Dear Reviewer,

Thank you again for your constructive feedback on our submission.

We would like to kindly follow up on our detailed responses and the revised manuscript, which includes substantial theoretical clarifications and additional experimental evaluations addressing your original concerns.

As the discussion period approaches its conclusion, we would sincerely appreciate your reconsideration of our responses and revisions. If any substantive concerns remain unresolved, we would be grateful for your further feedback.

Thank you very much for your time and consideration.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.