acceptodds
Under review as a conference paper at ICLR 2027

ShapFM: Amortized Shapley Estimation with a Prior-Fitted Model

14 views

Abstract

Shapley values are among the most widely used tools for explainable AI, with applications ranging from feature attribution and data valuation to causal explanation. In these settings, a machine learning problem is represented as a cooperative game over players, and the goal is to estimate each player’s Shapley value. Traditional estimators query each game and estimate Shapley values from scratch. In contrast, recent amortized approaches such as FastSHAP and ExplainerPFN learn across related games to predict Shapley values directly. These methods exploit shared structure, but their predictions are static: once trained, they cannot improve as additional evaluations of the current game become available. We introduce ShapFM, an adaptive amortized estimator that combines the strengths of both approaches. ShapFM is a prior-fitted transformer-based model, pre-trained on synthetic ML games to estimate Shapley values given evaluations of a game. Once trained, it yields Shapley estimations for real-world ML games with one forward pass. Like a traditional estimator, its predictions improve as more game evaluations are provided; unlike a traditional estimator, it can leverage structure learned from many related games to obtain accurate estimates from very few queries. We design the architecture specifically around the structure of cooperative games, alternating attention over players and coalitions, incorporating the underlying inputs that define each game, and optionally using evaluations of related games as additional context. Across a broad collection of machine learning games, ShapFM is competitive with state-of-the-art, heavily optimized traditional estimators and can substantially outperform them in the low-query regime.

Primary area:
interpretability and explainable AI
Keywords:
Shapley value, prior-fitted networks, amortized estimator, explainability, cooperative games

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_o8JROpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes ShapFM, an amortised method for estimating the Shapley values of cooperative games arising in machine learning. ShapFM is a prior-fitted, transformer-based model that handles games with up to 32 players. Unlike FastSHAP, which must be trained per model, ShapFM is trained once and can be reused across different games, models and datasets. Unlike ExplainerPFN, it conditions on evaluations of the game itself, so its estimates improve as more coalitions are observed.

ShapFM is trained only on about 960,000 synthetic games with exact Shapley values, drawn from four families: sums of unanimity games, Gaussian processes on the hypercube, TreeSHAP games of tree ensembles fitted on synthetic data, and do-Shapley games on linear causal models.

The input is a coalitions × players grid of observed coalition values, plus two optional inputs: related games and, for tabular data, the explicand with reference rows. One forward pass yields 999 quantiles per player, whose mean satisfies efficiency.

Experiments cover games with up to 24 players (tabular, image, text, hyperparameter importance and SOUM), against sampling-based estimators and the amortised explainers FastSHAP and ExplainerPFN.

Critical strengths
  • The framing is interesting: it bridges the foundation-model literature with amortised Shapley estimation, combining reuse across games with refinement from coalition evaluations, without per-model training.
  • A single checkpoint can potentially be applied to many cooperative games, not only feature attribution (e.g., hyperparameter importance and SOUM games).
  • The method is compared against recent baselines, including OddSHAP, RegressionMSR, ShaplEIG, FastSHAP and ExplainerPFN.
  • It outperforms the baselines in the low-budget regime (1.5 to 3.6× lower error at p + 2 evaluations on tabular games), which may matter in applications where evaluations are expensive.
  • Its inference runtime is comparable to that of existing estimators (53 ms without context, 123 ms with it) despite the larger context, and far below ShaplEIG's.
Critical weaknesses
  • The approach bets that real ML games resemble the synthetic, mostly low-order training games, leaving it open to "adversarial" games that behave very differently.
  • The evaluation is limited to games with at most 24 players (training to 32), well below many practical settings.
  • The advantage diminishes as the number of sampled coalitions grows, and classical estimators overtake it at larger budgets.
  • The method provides no theoretical bounds on its approximation error, unlike, e.g., LeverageSHAP.
  • Methods are compared only at an equal number of evaluations; a comparison at an equal time budget would be more informative.
  • The FastSHAP comparison may not be entirely fair: it uses a fixed 4,096 training rows, although FastSHAP's error should decrease with more training data, and its very cheap inference lets its training cost be amortised over many explanations.
  • The motivation is not entirely clear. It is unclear which real-world applications call for this trade-off: reasonably accurate explanations from very few evaluations, for models with few features, at the cost of relying on a learned prior. This reliance may be undesirable in high-stakes domains such as credit decisioning, where explanations should be faithful to the model rather than shaped by a synthetic prior.
  • The uncertainty represented by the 999 output quantiles is never analysed (e.g., calibration), although it is presented as a feature of the method.
Questions for authors
  • Regarding the Gaussian process games, the paper states: "The family has no dummy players and no symmetric pair, and all distributions were calibrated on a small set of real precomputed training games." What are these games? If they overlap with the datasets used for evaluation, there could be leakage. How does ShapFM perform without this leakage, if any?
  • Could the authors provide more details on the training cost of FastSHAP vs. ShapFM: training time, GPU memory requirements, which GPU was used, etc.?
  • Why are no ShaplEIG results reported for the tabular games in Table 1, Figure 4 and elsewhere in the paper? ShaplEIG appears to be the closest competitor (modulo its high computational cost), so it would be valuable to see how it compares with the proposed method on tabular games.
  • Beyond the runtime at a fixed evaluation budget, could the authors report the approximation error at a fixed runtime budget, to better compare estimators? Can more light be shed on the training cost of FastSHAP compared to ShapFM?
  • ShapFM should work well when the fitted prior (mostly low-order interaction games) matches the game at hand. What happens when it does not?
  • Could the authors further motivate which applications would benefit from the method? It is not immediately clear where obtaining amortised Shapley values from very few sampled coalitions with a prior-fitted model is beneficial in real-world applications.
  • Connected to the above. Can anything be said about the theoretical guarantees the proposed framework provides?
  • How well do the 999 quantiles capture the uncertainty of the Shapley values? Does the predicted distribution widen when an unusual game is provided, either an unusual sample within a dataset or a dataset that is unusual with respect to the learnt prior?
6 comments

We thank you for the detailed and constructive feedback and for your positive assessment of our work. Below, we respond to each question and weakness point by point. Several of your comments call for additional experiments, and we will upload a revised manuscript with these results by the end of the week. Where a response depends on these new results, we describe the planned changes here and will follow up once the revision is online. If you have any follow-up questions or suggestions in the meantime, please let us know so that we can take them into account for the revision!

Q1: Regarding the Gaussian process games, the paper states: "The family has no dummy players and no symmetric pair, and all distributions were calibrated on a small set of real precomputed training games." What are these games? If they overlap with the datasets used for evaluation, there could be leakage. How does ShapFM perform without this leakage, if any?

The GP prior was calibrated on a subset of precomputed games available in shapiq [1], ranging from local explanations to data valuation (always the first instance (seed 1) and restricted to at most 15 players). These games are not the same as the ones used for the evaluation of ShapFM, and hence there is no risk of leakage.

Moreover, we only conduct prior predictive checks, i.e., compute summary statistics per game (e.g., sign share, size-trend) and adjust the prior’s hyperparameters until the real games’ statistics fall inside the 5-95% band of the simulated distribution. So only these statistics enter the procedure, never coalition or Shapley values.

Q2: Could the authors provide more details on the training cost of FastSHAP vs. ShapFM: training time, GPU memory requirements, which GPU was used, etc.?

That’s an important question! We agree that the training cost belongs in the paper and will add details to the revision.

The 32-player ShapFM model was trained on a single NVIDIA L40S GPU with 48 GB of memory. It was pretrained once, which took 11 hours and 14 minutes. The larger 128-player ShapFM model was trained for about 7 days on two NVIDIA A100 GPUs with 40 GB each, followed by a final phase of 16 hours on one NVIDIA B300 GPU, which corresponds to about 30 hours on an A100. In total, this is about 350 A100 GPU-hours, with about 27 GB of memory used per GPU.

A direct comparison with FastSHAP is difficult, because the two methods require training for different things. ShapFM is pretrained once and then explains any game without further training. FastSHAP must be trained anew for every game it explains, and its cost grows with that model. On ImageNette with 196 players, reported training times are about 2.4 hours on one Tesla V100 [2], about 19 hours on two RTX 2080 Ti [3] and about 4 hours on one RTX 2080 Ti [4].

Q3: Why are no ShaplEIG results reported for the tabular games in Table 1, Figure 4 and elsewhere in the paper? ShaplEIG appears to be the closest competitor (modulo its high computational cost), so it would be valuable to see how it compares with the proposed method on tabular games.

Yes, of course! As you mention, ShaplEIG is a very slow but powerful method, and we will report ShapEIG for all experiments at the budgets p+2p+2, 2p2p and 4p4p (Table 1, 11 and 12). Due to its computational cost, we cannot cover the full budget range, since a single ShaplEIG explanation of a game with 24 players at a budget of 16p16p costs about 6 CPU hours, or about 180 CPU hours for all 30 tasks, and the cost grows further up to the largest budget of 32p32p.

Q4: Beyond the runtime at a fixed evaluation budget, could the authors report the approximation error at a fixed runtime budget, to better compare estimators? Can more light be shed on the training cost of FastSHAP compared to ShapFM?

To make sure we understand this question correctly, we kindly ask for a short clarification. Figure 7 reports the time per explanation at a fixed budget of 8p evaluations. Do you mean the reverse view, where every estimator receives the same amount of time and uses as many evaluations as fit into it?

Q5: ShapFM should work well when the fitted prior (mostly low-order interaction games) matches the game at hand. What happens when it does not?

Our evaluation already covers this case, since many of the benchmark games differ substantially from the prior. No game in our evaluation was seen in training. Most also differ from the prior in distribution, as we detail below. The two “in-prior” families are less in-prior than they appear. The tree family of the prior is fitted on synthetic data only, and the evaluation uses real datasets only (Section E.2). The sum-of-unanimity (SOUM) games draw 30 or 150 terms of order up to 5 or 15, whereas the prior draws at most 12 terms per block of order at most 6 (Table 3). Thus, the SOUM games lie well outside of what is seen during training. Still, ShapFM has the lowest error on them at 2p (NMSE 0.390 and 0.462 vs. 0.476 and 0.540 for the best estimator, Table 11).

Several evaluation families have no counterpart in the prior at all. The deep learning games already differ from the prior quite substantially in the model class (ResNet, ViT, DINOv2, DistilBERT). Of these six game families, ShapFM is still the most accurate estimator at budget 2p and at 4p on five of six, where only ShaplEIG is ahead on ResNet-18 (Table 11), but with drastically higher computation time. We will further strengthen confidence in our model's generalization capability with more experiments on these deep learning explainability games.

Q6: Could the authors further motivate which applications would benefit from the method? It is not immediately clear where obtaining amortised Shapley values from very few sampled coalitions with a prior-fitted model is beneficial in real-world applications.

Indeed, ShapFM’s main strength is data-scarce applications where no closed-form solution is available. We see our approach as particularly advantageous when observing coalitions is costly and we have access to context such as related games.

Three concrete examples, which we also study in the paper, are: (1) feature importance for ML pipelines (where obtaining coalitions requires refitting the whole pipeline), (2) hyperparameter explanations (where obtaining coalitions requires running complete hyperparameter runs), and of course (3) local explanations for black-box models such as image classifiers, language models, or tabular foundation models (where we often have access to prior explanation games on similar images / text documents / rows).

More generally, we expect that all application areas where Shapley-based explanations are needed for black-box models will benefit from a well-trained foundation explanation model such as our ShapFM. Ultimately, the goal is to have practically useful Shapley values. Meaning (a) a high-quality estimate of the Shapley values and (b) with the least amount of black-box access (since this is the thing that is costly).

Even with advances such as ShaplEIG or other proxy-based techniques like RegressionMSR or OddSHAP, traditional estimators only work well given sufficient sampling budget, as they cannot easily reuse information about the underlying explanation task or machine learning model. The answer, up to now, was to build a FastSHAP model for each individual explanation task from scratch. While technically you would no longer need any black-box access, training a good FastSHAP model is quite finicky, and practical guidance seems to only really exist for local explanations of vision models and some tabular cases.

Foundation models, such as ShapFM, bridge this gap now. Similar to FastSHAP, we can now train the Shapley explanation model on synthetic data and enrich it with as much additional context as possible such that the model already works with very little information, as you can see from our context vs. w/o context ablations (e.g., Figures 3 and 4). Unlike FastSHAP, the performance of the explanations can be assured by a central model provider training on various different game settings and scenarios and by providing additional queries to the black box, reducing the estimation error like a traditional sampling-based approximation technique.

Q7: Connected to the above. Can anything be said about the theoretical guarantees the proposed framework provides?

This is an important point we have talked about in different parts of the paper. Theoretically, ShapFM provides guarantees at two levels.

By construction, for every game and budget, it satisfies efficiency exactly (Eq. 4, Sec. 2.1). As an orthogonal projection onto the efficient hyperplane, Eq. 4 can only reduce the squared error (cf. Jethani et al., 2022, App. B). ShapFM is equivariant to player order (Sec. 2.1, B.1), and its outputs transform correctly under shifts and positive rescalings of the game (Sec. D.1). Under the conditions discussed in the general comment (C2), the optimal solution of our training objective (Eq. 12) is the quantiles of ϕ\phi given the observations and context under the training distribution, and our point estimate, the mean of these quantiles, approximates the conditional mean of ϕ\phi. As with FastSHAP, whose guarantees also hold relative to its training objective ([2], [3]), this gives no per-game bound.

Importantly, ShapFM is neither unbiased nor provably consistent, dummy and symmetry are only learned through the prior (Sec. C.1), and linearity is not enforced. Classical estimators, in contrast, are consistent, so they are expected to overtake ShapFM at large budgets, as our results show (Sec. 3.2, Figs 5 and 8). This is a deliberate trade-off, as ShapFM targets the low-budget regime.

Q8: How well do the 999 quantiles capture the uncertainty of the Shapley values? Does the predicted distribution widen when an unusual game is provided, either an unusual sample within a dataset or a dataset that is unusual with respect to the learnt prior?

We agree that understanding (and using) these uncertainties offers great opportunities for future research, particularly for uncertainty quantification and calibration behavior. This — from our point of view — is still quite underexplored, also for tabular foundation models.

For the revision we will add a calibration study where we will compare the predicted quantiles with the true Shapley values on a subset of our benchmark games.

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

This paper studies the problem of amortized Shapley value estimation for cooperative games, where estimating Shapley values typically requires a large number of coalition evaluations. The authors propose ShapFM, a prior-fitted transformer pretrained on synthetic games drawn from four families with exactly computable Shapley values. The proposed architecture alternates attention across coalitions and players to preserve permutation equivariance, and further supports incorporating contextual information from related games and underlying feature data. Once pretrained, ShapFM can directly estimate Shapley values for unseen games from a limited number of coalition evaluations without task-specific training. The authors evaluate the proposed method across several machine learning tasks, including feature attribution, hyperparameter importance, and synthetic games. Experimental results demonstrate competitive performance against existing Shapley value estimators, particularly in the low-query regime.

Critical strengths
  1. The proposed architecture is well motivated by the structure of cooperative games. By alternating attention across coalitions and players, the model preserves permutation equivariance with respect to players and can naturally handle varying numbers of players and coalitions. Moreover, the architecture allows additional contextual information from related games to be incorporated, providing the potential for more accurate Shapley value estimation through contextual learning.
  2. The authors construct a diverse synthetic training prior consisting of four game families with exactly computable Shapley values. This is a reasonable and practical approach to obtaining large-scale training data without expensive Shapley value estimation. The resulting model demonstrates promising performance across several downstream tasks, particularly in the low-query regime.
Critical weaknesses
  1. The generalization capability of the proposed method remains insufficiently investigated. Although the paper evaluates ShapFM across different machine learning tasks, several evaluation settings share substantial structural similarities with the synthetic training prior. For example, both training and evaluation include tree-based explanation games and sum-of-unanimity games. While the image and text experiments provide some evidence of generalization beyond the training families, it remains unclear how well ShapFM performs when the underlying game structure differs significantly from the training distribution. More controlled out-of-distribution experiments would help establish the generality of the proposed approach.
  2. The contribution of each synthetic prior family is not sufficiently analyzed. The authors combine four different game families to pretrain ShapFM, but the paper does not provide ablation studies examining the necessity or individual contribution of these families. For example, it would be interesting to investigate whether a model trained on only one or two families could achieve comparable performance, and how removing a particular family affects generalization to different downstream tasks. Such experiments would provide a better understanding of what the model learns during pretraining and which aspects of the synthetic prior are essential for its effectiveness.
Questions for authors

Please see weaknesses.

3 comments

Thank you for your thoughtful review and for recognizing that our architecture is well motivated by the structure of cooperative games and that our synthetic prior is a practical way to obtain large-scale training data. Since your questions refer to the weaknesses you raised, we address them jointly below. For both points, we will provide additional results and include them in a revised manuscript by the end of this week. Until then, we describe the planned changes here and will follow up once the revision is online. If you have any further thoughts or suggestions in the meantime, we'd be very happy to hear them and take them into account!

W1: The generalization capability of the proposed method remains insufficiently investigated. [...] several evaluation settings share substantial structural similarities with the synthetic training prior. For example, both training and evaluation include tree-based explanation games and sum-of-unanimity games. [...] it remains unclear how well ShapFM performs when the underlying game structure differs significantly from the training distribution. More controlled out-of-distribution experiments would help establish the generality of the proposed approach.

We agree that controlled “out-of-distribution” experiments would strengthen the paper and we will provide them in the revised manuscript in the form of prior ablation. See also our response to W2 and question Q5 to reviewer o8JR.

W2: The contribution of each synthetic prior family is not sufficiently analyzed. [...] the paper does not provide ablation studies examining the necessity or individual contribution of these families. For example, it would be interesting to investigate whether a model trained on only one or two families could achieve comparable performance, and how removing a particular family affects generalization to different downstream tasks. [...]

Thank you for this suggestion! Please also see our general comment (C1) on upscaling the model, as we have retrained ShapFM for higher-player games. We have retrained the 128-player variant on reduced priors and report the NMSE relative to the full four-family mixture. We will add this as an additional ablation in the revision. The summary from this ablation is that no reduced prior beats the full mixture overall, meaning the full set of priors leads to the best ShapFM performance. Removing the Gaussian process prior leads to a decrease in performance on games with non-structured data like text or images, while the other families are essential for games related to tabular data. Surprisingly, removing the prior on causal models has only a marginal effect on tabular games. Keeping it provides a small but consistent gain on images and text, indicating that the SCM-based prior contributes a complementary inductive bias beyond tabular structure. We thank the reviewer for raising this question, as it clarified the specific role of the SCM-based prior in the mixture. We had expected it to help primarily on structured data, since SCMs are the standard pretraining prior for tabular foundation models, yet its contribution lies mostly in unstructured modalities.

Thanks for the rebuttal, which has addressed most of my concerns. I will maintain the positive score.

We thank the reviewer for the valuable feedback!

Reviewer_eaBtOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Research paper proposes ShapFM - pretrained transformer based on synthetic cooperative games with exact Shapley values. Using evaluations of new games it predicts Shapley value vector in one forward pass and estimates the improvement based on additional evaluations. ShapFM can use simila game models and input data as extra context. Numerical experiments are based on around 50 game families and compare ShapFM with sample-based estimators and methods such as FastSHAP.

Critical strengths
  1. Interesting idea. Training one model to estimate Shapley value vector across many games and refining estimation with new evaluations of current game. This is an interesting way to combine amortization with classical estimation of Shapley value.

  2. Based on fundamental applied mathematica - game theory. Model design respects key Shapley value properties such as model treats players equally and satisfies efficiency by design. Taining dataset of game models contains with exact Shapley values.

  3. Research paper is well structured and written. Research paper is clear and easy to follow, numerical experiments cover wide range of game families.

Critical weaknesses
  1. Problem statement is not new. Shapley value vector estimation is mature area with wide range of existing sample-based and amortized methods. Even authors present ShapFM as combination of two existing methods.

  2. Model architecture is borrowed. Alternating attention over rows and columns follows TabPFN and TabICL. Finally main new part is synthetic games models dataset.

  3. Dataset is limited for small games. Model is pretrained on game models with up to 32 players. The advantage of this model is deacreasing as budget grows. Many real attribution tasks have more features than that.

  4. No convergence guarantee. Research paper gives no guarantee that the method approaches exact Shapley value vector as more evaluations are provided. Accuracy depends on how close target game is to synthetic training dataset.

Questions for authors
  1. How ShapFM behaves on game models with > 32 players involved?

  2. Shapley value vector estimation converges to the exact values when all coalitions are given?

4 comments

Thank you for the detailed and constructive feedback on our work. Below, we respond to each question and weakness point-by-point. Your question about games with more than 32 players was especially valuable, as we train and evaluate ShapFM with up to 128 players for the rebuttal, where its advantage over existing estimators grows substantially in the low-budget regime. We will upload a revised manuscript with these and further results by the end of the week. Where a response depends on the new results, we describe the planned changes here and will follow up once the revision is online. If you have any follow-up questions or suggestions in the meantime, please let us know, so that we can take them into account! We hope that our responses and new results address your concerns.

Q1: How ShapFM behaves on game models with > 32 players involved?

This is, in our view, the most important point, which was also raised by the remaining reviewers. This is why we have conducted additional training and evaluation runs, training ShapFM with up to 128 players.

In short, we expect ShapFM to follow the scaling behavior of ML models; i.e., scaling up ShapFM requires larger synthetic games and longer pre-training (and eventually also a higher parameter count to increase capacity). Accordingly, the 128-player ShapFM model keeps the architecture, loss and optimizer and differs only in larger synthetic games, re-tuned hyperparameters and longer pre-training. Please also see the general comment C1 for more information on this.

However, the scaled-up variant already reveals a great benefit of our approach in that the ShapFM improves even more compared to existing baselines, particularly in the low-budget regime we are targeting here. Classical, SOTA estimators such as LerverageSHAP require pp or 2p2p coalitions to be properly defined. ShapFM already retrieves high-quality estimates with much fewer coalitions. Evaluating ShapFM on the existing evaluation games shows that we retain the performance of the 32-player ShapFM variant. Evaluating ShapFM on larger-pp games shows that the existing baselines do not return practically meaningful estimates. This becomes especially apparent when comparing estimators with the precision metric which shows that the 128-pp ShapFM performs substantially better in identifying top-k players than all existing baselines.

Q2: Shapley value vector estimation converges to the exact values when all coalitions are given?

Neither by construction nor by the pre-training setting we used; however, this is also not the regime ShapFM is designed for. When all coalitions can be evaluated, Eq. 1 gives the exact Shapley values directly, and no estimator is needed. In this case, the context contains the complete game, so the Context Error vanishes, and ShapFM's error depends only on the Prior Error and the Amortization Error (see our response to your W4 and C2 in the general comment). In addition, to focus pre-training on small budgets, we cap the number of coalitions in the context at min⁡(512,(2p−2)/2)\min(512, (2^p-2)/2), so the model never sees a complete game during pre-training. Thus, in practice, exact recovery is not guaranteed. To illustrate this, we will compute the NMSE of ShapFM on small-scale games with all coalitions provided in the context and add this to the appendix.

W1: Problem statement is not new. Shapley value vector estimation is mature area with wide range of existing sample-based and amortized methods. Even authors present ShapFM as combination of two existing methods.

Exactly! Shapley estimation is a very active and important research field. This reflects how important the problem is for downstream applications. Our contribution is not the problem itself but a novel way of solving it. To the best of our knowledge, ShapFM is the first estimator that learns across games and still improves as more queries of the current game become available.

ShapFM brings together the strengths of two families of approaches whose limitations are complementary (Figure 1, left). Sample-based estimators start from scratch for every game and ignore related games and the predictor's data. FastSHAP explains an input in a single forward pass, but it has to be retrained for every explanation task, and its estimates cannot be refined with further queries. ShapFM is pretrained only once, and we apply a single checkpoint of the model to every game and budget (Section 3). Achieving this takes more than putting existing methods together, since it depends on a prior of games with exact Shapley values (Section 2.3) and on the game-specific architecture we describe in our response to W2.

This pays off when black-box access is scarce (Table 1, Figure 4). Context alone, which sample-based estimators do not use, lowers the error by up to 2.2x (Figure 3). We also see ShapFM as a first instance of this paradigm, and in future iterations, better priors can directly lead to better ShapFM-like approaches.

W2: Model architecture is borrowed. Alternating attention over rows and columns follows TabPFN and TabICL. Finally main new part is synthetic games models dataset.

Yes, attention over two axes and induced set attention are established building blocks in SOTA tabular foundation models like TabPFN, TabICL, and the Set Transformer (as we acknowledge in the paper).

In contrast, our learning task is completely different from what tabular foundation models learn. TabPFN and TabICL learn to predict labels (for a query) from labeled examples (context), i.e., the data structure of context and query is the same. In our setting there is no dedicated query, and ShapFM rather learns to approximate a mathematical operator on cooperative games given context (Section 1.2). Its cell encoding selects separate value maps for players inside and outside a coalition (Eq. 3), following the coalition-value decomposition of the Shapley value used by SVARM (Eq. 5, Kolpaczki et al., 2024). Cross-attention connects each cell to the same player in related games (Section 2.2), and a natural-context encoder reads the explained model's data. We also adopt the additive efficient normalization used by FastSHAP so that efficiency holds by construction (Eq. 4).

Thus, we respectfully disagree that the synthetic games are the only new part. However, we do see the prior as a core contribution, as it is for prior-fitted models in general. It provides exact Shapley targets across four game families and organizes them into groups of related games (Section 2.3).

In our analysis, we show that these choices matter empirically (Section 3.3, Figure 6). Removing all context raises the error by up to 2.4x, with related games accounting for most of this rise. TabPFN v2 alternates between its two axes as we do, whereas TabICL runs column attention before row attention, but such a sequential design raises our error by up to 2x. A flat cell encoding raises it by up to 1.2x.

W3: Dataset is limited for small games. Model is pretrained on game models with up to 32 players. The advantage of this model is deacreasing as budget grows. Many real attribution tasks have more features than that.

We agree, and we acknowledge this limitation in Section 5. We want to clarify that nothing in our architecture ties ShapFM to the 32 players. This is why we can up-scale ShapFM to, for example, 128 players, which we will add to the revised manuscript. For more information, see our response to your question Q1 and comment C1 in the general comment.

W4: No convergence guarantee. Research paper gives no guarantee that the method approaches exact Shapley value vector as more evaluations are provided. Accuracy depends on how close target game is to synthetic training dataset.

This is correct. As described in C2 in the general comment, the Prior Error and the Amortization Error do not vanish with more evaluations, even when all the coalitions are evaluated (see our response to your question Q2). Thus, unlike Monte Carlo or regression estimators, ShapFM has no asymptotic consistency guarantee, similar to FastSHAP or ExplainerPFN. We can, as you mentioned as a critical strength, guarantee exact efficiency for every game and budget (Eq. 4) and equivariance to player order.

Reviewer_eaBtOpenReview ↗

Comment for "Official Comment by Authors"

Thank you for the clarrifications. Since the training and architectural part of the paper is directly connected to how The Shapley Value is computed and numerical results are good, I will increase the score.

We thank the reviewer for the constructive feedback and for raising the score!

Comments on the paper

3 comments

First and foremost, we sincerely thank all reviewers for their time and their thoughtful feedback on our work. Many of your comments are best addressed with additional experiments and ablations, which we are currently finalizing. We will update the manuscript with these results towards the end of the week and upload a changelog alongside it.

We know that reviewing is a considerable effort, and we do not want to put any pressure on you to respond quickly. We are posting our responses ahead of the revision so that any follow-up points can still inform it, and we would be very grateful for any further feedback you are able to share.

C1 Scaling up: [eaBt, QtKU, o8JR]. All reviewers point out the limited player count as a central limitation of our current manuscript (stated as the main limitation in Section 5). We agree and see this as the most important issue for improving the manuscript. Since nothing in our architecture ties ShapFM to 32 players, we increase the player count from 32 players to 128 players and train a 128-player variant of ShapFM. We train the 128-player ShapFM model on the same four prior families as described in the paper, but sample larger games from them. We do not change anything in the underlying model architecture.

We will make small adjustments to the updated manuscript and (i) add the new 128-player variant to the existing evaluation pipeline as a new approximation curve, (ii) compare it with the existing 32-player variant in a separate ablation, and (iii) add games up to 121 players to the existing approximation curves. Since ground-truth comparisons are not possible with black-box functions that have a computationally prohibitive number of players (p>25p > 25), we compute “ground-truth” scores via LeverageSHAP and a sampling budget of 16 38416\,384 coalitions.

We evaluate the 128-player variant on a broader range of games up to 121 players, including larger variants of deep learning explanation games (e.g., DINOv2, ResNet or TabPFNv3). Comparing the performance of the 128-player ShapFM variant shows that it substantially improves over the traditional approximation methods in the low-budget regime, illustrating the new paradigm of utilizing prior-fitted models for Shapley estimation. Many SOTA approximation methods, such as LeverageSHAP or OddSHAP, require pp or 2p2p evaluations of the underlying game to be properly defined. ShapFM already retrieves high-quality estimates for budgets way below that. Further, in addition to the NMSE (normalized mean squared error), we also evaluate the methods in terms of Precision (Precision@k measures the performance in detecting the top-k most important players in terms of absolute scores), which shows that the estimates remain meaningful at scale.

As suggested by [QtKU], we conduct an ablation of the four prior families with this scaled-up 128-player ShapFM variant. We have retrained the 128-player variant on reduced priors and report the NMSE relative to the full four-family mixture. We will add this as an additional ablation in the revision. The summary from this ablation is that no reduced prior beats the full mixture overall, meaning the full set of priors leads to the best ShapFM performance.

C2 In-Prior vs. Out of Prior: All reviewers ask about the dependence on the training prior or convergence guarantees. As these are closely connected, we want to first clarify what kind of estimator ShapFM is. ShapFM is trained completely on synthetic games and does not see any real games during training. Unlike a classical estimator, which applies a fixed formula to one game, it learns a mapping from a few observations to plausible Shapley values, combining a prior over games with the current game's observations.

Although not exactly a prior-data fitted network (see our discussion in the paper; l156-160), ShapFM is pre-trained on prior-generated data with a pinball loss, and is therefore trained to estimate quantiles of the Shapley values conditional on the context, i.e., it approximates Shapley values. In practice, the error of this approximation consists of (a) the Prior Error, which describes the mismatch between the synthetic games and real games, (b) the Amortization Error, which depends on how well the model is trained, and (c) the Context Error, which characterizes the lack of information in the context [1].

This means that, for the two points mentioned in the reviews: First, under specific conditions, our setup might converge to the optimal solution; however, in practice, we do not know whether these specific conditions apply. Second, more coalitions in the context will reduce the context error, but not the other two errors; hence, we cannot claim convergence or good out-of-prior estimation and must rely on thorough empirical analysis (as we do in the paper). Finally, to relate this to other estimators: Like FastSHAP and ExplainerPFN, ShapFM carries an approximation error from training that cannot be driven to zero by more data at inference time. This matches our results, since classical estimators overtake ShapFM at large budgets (see Section 3.2, Figures 5 and 8).

Ultimately, the prior should cover all relevant games we expect to encounter, so we do not aim to perform well on games that are completely out of prior. We discuss this in more detail in answer Q5 to reviewer o8JR, where we also explain what out-of-distribution tasks we actually use in the experiments.

[1] Hülskamp et al. How Well Do Prior-Data Fitted Networks Amortize Bayesian Inference? In: arXive 2610.13764

We thank all reviewers again for their time and constructive feedback, which helped us considerably improve the paper. We have uploaded a revised version with all changes highlighted in blue\color{blue}{\text{blue}}. Below, we briefly summarize the main changes.

Changes

  • Added ShapFM (128p), pretrained on up to 128 players (Section 2.3, App. C and App. D.2). Compared to the 32-player ShapFM, we only change a few hyperparameters and prior settings (Tables 7 and 10). Both models perform similarly (Tables 12 and 13), so the main paper now uses the 128p model. We are porting all ablations to it for the camera-ready version and keep a table showing how the 32p model scales to more players (Table 20). [eaBt, o8JR]
  • Extended the evaluation benchmark to games with up to 127 players (Section 3, App. E.2; overview of the benchmark suite in Table 11). Above 24 players, we use a LeverageSHAP estimate as ground truth and additionally report Precision@k. On large TabPFN games (36 to 111 players), ShapFM (128p) achieves a mean NMSE of 0.088 at 2p, compared to 0.275 for ShaplEIG. [eaBt, o8JR, QtKU]
  • Updated the results with ShaplEIG scores for almost all games in the benchmark suite. ShaplEIG results are still missing for the interventional games on the MIC dataset (RF and XGBoost); we will report them once they are available. With context, ShapFM is 1.6 to 2.4 times more accurate than the best competing estimator on the tabular game families. [eaBt, o8JR]
  • Revised the comparison of ShapFM with FastSHAP and ExplainerPFN (Section 3.2, App. E.4, Table 18). Students trained on ShapFM estimates are more accurate on the random-forest games but less accurate on the MLP games. We have therefore withdrawn the claim that ShapFM provides a better training signal. App. E.4 (Figure 15, Table 19) trains FastSHAP on 256 rows up to the full training set and reports its training cost. With all rows, FastSHAP beats ShapFM at a budget of 32 in three of four cases. [o8JR]
  • Added an ablation of ShapFM's prior (Section 3.5, App. E.6, Figures 8, 19 and 20). Removing unanimity games increases the error on tabular games by up to 1.41×, and removing TabICL by up to 2.12× (with context). Removing the Gaussian process increases the error by 1.28× on the image game but decreases it on tabular games. [QtKU]
  • Added an evaluation of ShapFM's calibration (App. E.7, Figure 22). The intervals are overconfident at small budgets and must be recalibrated before use. Uncertainty decreases as more coalitions are evaluated (Figure 22(c)). [o8JR]
  • Added plots of error against runtime on 12 pre-computed benchmark games, on both CPU and GPU (Figures 22 and 23). Figure 7 now includes ShapFM's GPU and CPU runtimes. [o8JR]
  • Added App. E.9, which studies the out-of-prior behavior of ShapFM in three settings: exceeding the player range of ShapFM (32p), increasing the interaction size beyond the prior, and evaluating ShapFM with all coalitions. [eaBt, o8JR, QtKU]
  • Added the training costs of ShapFM (App. D.1). ShapFM (128p) took about 350 A100 GPU hours; ShapFM (32p) took 11 hours and 14 minutes on an NVIDIA L40S. [o8JR]

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.