acceptodds
Under review as a conference paper at ICLR 2027

Tensor Network Datamodels for Training Data Attribution

Abstract

Datamodels predict how a model’s behavior changes when different parts of its training data are included or removed. However, linear datamodels assign fixed contributions to training units and cannot represent non-additive joint effects. We introduce Tensor Network Datamodels (TNDMs), which use low-rank tensor networks to model dependencies among training units while keeping the representation compact. TNDMs learn a map from subsets of the training data to the outcome of training on those subsets, allowing them to represent interactions among several training units. We prove that complementary and redundant training groups can induce high-degree subset functions with compact low-rank representations. A single fitted TNDM can then be used for several training data questions, including predicting the effect of removing data, measuring the importance of individual training units, identifying interactions between training units, and selecting useful subsets of the training data. We evaluate TNDMs on controlled tasks designed to contain strong interactions, as well as standard fine-tuning settings for classification and reasoning. On a controlled reasoning task, TNDMs outperform every baseline while observing only half as many training subsets, and TNDM-TT reduces prediction NMSE by relative to the linear datamodel after observing approximately of all possible training subsets. In factual-association fine-tuning, a single TNDM-CP raises held-out from to over the linear datamodel. These results demonstrate that compact low-rank models improve prediction when training data effects depend on combinations of training groups.

Primary area:
interpretability and explainable AI
Keywords:
Training-data attribution, datamodels, tensor networks, low-rank approximation, higher-order interactions, cooperative game theory, data valuation.

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_U1ZWOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces Tensor Network Datamodels (TNDMs), using CP and TT representations to predict model outcomes from training-data subsets. Its central motivation is that complementary or redundant training groups can induce high-degree interactions while retaining low tensor rank. Experiments evaluate subset prediction, attribution, interaction recovery, and data selection on controlled tasks and pretrained-model fine-tuning.

Critical strengths
  1. The paper addresses a meaningful limitation of additive training-data attribution: the utility of a training group can depend on which other groups are present. The distinction between interaction degree and tensor rank provides a useful motivation for modeling these dependencies compactly.
  2. The controlled experiments provide encouraging evidence, including improved subset prediction and interaction recovery. Appendix J.15 is particularly interesting: interactions change from synergy to redundancy during training, while non-additive effects remain compressible. These observations suggest useful directions for studying training-data dependencies.
Critical weaknesses

The evidence connecting the proposed structural explanation to practical benefits on natural data remains incomplete. The paper acknowledges its relationship to TN-SHAP and established tensor representations, making the training-data-specific insights particularly important. Sport–Music shows improved prediction, but does not clearly establish a distinctive higher-order interaction benefit over quadratic models. On AG News, cross-budget comparisons favor degree-truncated and RBF baselines. Moreover, the dynamics in Appendix J.15 come from exhaustive enumeration on a controlled task, rather than recovery by a sparsely fitted TNDM. Thus, the paper does not yet establish when the proposed structure reliably occurs and translates into better data interventions.

Questions for authors

What evidence supports the connection between higher-order, low-rank structure and the observed gains in the fine-tuning experiments? In particular, do existing results demonstrate improved joint-removal prediction or data selection against strong nonlinear baselines on natural data? Please distinguish demonstrated results from proposed future experiments.

2 comments

We thank the reviewer. The revision addresses expressivity, spurious interactions, and training order.

1. Definition 3 and prior work. We agree that fitting Definition 3 is a standard low-rank tensor completion problem. Our contribution is a principled, unified formulation of datamodeling that makes degree bounds, sparsity, and low rank explicit (Section 2), and an analysis of why low rank suits training-data dependencies (Section 3): complementary and interchangeable groups can generate high-order interactions with compact tensor representations. We show that settings where the data consists of complementary and redundant groups is where linear datamodels fail, where TNDM represent the training data model compactly. Appendix A.6 makes the distinction from sparsity precise. An expansion with ss nonzero coefficients in any product basis has CP rank at most ss. Our coverage construction with mm required relations has TT ranks at most two under a paired ordering, but CP rank 2m2^m, requiring at least 2m2^m coefficients in every exact product-basis expansion (Theorems 5–6). Section 4.4 and Table 23 connect this structural motivation to the eight-hop transformer experiment.

TN-SHAP and TN-SHAP-G [1,2] are specifically focused on Shapley values and interactions and explain predictions from a fixed model. TNDMs predict outcomes after training on unseen data subsets. Park et al. [3] develop tensor-product models for functional ANOVA. We connect training-group dependencies to tensor rank and test prediction, joint removals, and attribution, building on [1,2] for attribution computations (Sections 2, 3, and 5).

2. Additive expressivity and spurious interactions. We agree that an additive function with d nonzero effects can require CP rank d, whereas TT represents it with ranks at most two. Appendix K.3 tests whether TNDMs assign interactions to noninteracting groups on the Figure 4 retraining game, where 22 of 28 pairwise and 52 of 56 third-order interactions are zero. At 64 queries, TNDM (CP) has the smallest spurious interactions (Section 4.4; Appendix K.3, Table 27) among the compared methods: normalized RMS 0.092 for pairs and 0.161 for triples, versus 0.451 and 0.461 for KernelSHAP-IQ. Its false-positive rates are 2% and 6%, versus 29% and 27%. TNDM (TT) gives 0.189 and 0.239 (Table 27).

3. Model scale and training regimes. Retraining cost, not the datamodel, limits the number of groups: each observed subset requires one training run. The fitted TNDM grows linearly with the number of groups, and the 24-group reasoning experiment (Section 4.1; Table 1) predicts outcomes over 224 subsets from at most 1,024 observed runs. On Qwen2-7B fine-tuning (Section 4.3; Tables 2 and 23) with ten optional groups and fixed background data, CP improves held-out R2 from 0.565 for quadratic regression to 0.668, and removing terms above degree two lowers it to 0.634 (Appendix K.2, Table 23).

4. Training order. A datamodel describes outcomes under a fixed training procedure, including example order, or averages over that randomness. We add a controlled experiment to training order ablations and a dedicated paragraph at the of the Appendix K.3. In a batch-order control, reversing or permuting the order changes negative held-out cross-entropy by 0.0033 and 0.0021 on average (Appendix K.3). AG News uses seeded shuffling; the Figure 4 learner uses full-batch training, which has no example order.

References

[1] Farzaneh Heidari, Chao Li, and Guillaume Rabusseau. Tractable Shapley Values and Interactions via Tensor Networks. AISTATS, PMLR 300:3043–3051, 2026.

[2] Farzaneh Heidari and Guillaume Rabusseau. TN-SHAP-G: Graph-Structured Tensor Network Surrogates for Shapley Values and Interactions. ICML, PMLR 306, 2026.

[3] Seokhun Park, Insung Kong, Yongchan Choi, Chanmoo Park, and Yongdai Kim. Tensor Product Neural Networks for Functional ANOVA Model. ICML, PMLR 267:48041–48085, 2025.

The comparisons show that no single method is best across all evaluated settings. Our contribution identifies a complementary regime where training-data dependencies admit compact low-rank representations despite requiring many interaction terms in a product-basis expansion (Section 3; Appendix A.6). The eight-hop transformer experiment illustrates the value of this structure (Section 4.4; Table 23), while AG News classifier refitting demonstrates improved joint-removal prediction (Section 4.3; Table 25). Figure 4 additionally shows that CP provides accurate individual and higher-order attributions from one fit, with errors competitive with ProxySPEX at 64 observations.

Thank you for the helpful review. Changes are in blue in the revised PDF.

  • When the structure appears: Appendix A.6 shows that complementary and interchangeable groups give TT ranks at most two but exponentially many coefficients in sparse bases.
  • Effect on interventions: the four-part relation learner has this setup. TT lowers prediction/removal error from 0.251/0.131 (ProxySPEX tree) to 0.0046/0.0061 and recovers Shapley values and interactions more accurately in all three learner seeds (Table 24).
  • Natural data: on AG News classifier refitting, TT has the lowest two-group removal error at 256 and 512 observations (0.144 and 0.094 vs. 0.575 and 0.524 for ProxySPEX; Table 25), and the lowest selection regret at sizes three and five on RoBERTa–LoRA (Table 26).
  • Higher-order terms: removing CP terms above degree two lowers Sport–Music R2R^2 from 0.668 to 0.634 (Table 23).

We hope to have addressed your concerns and answered your questions. We remain available if you have any more questions during the end of the discussion period.

If we properly addressed your concerns and questions, we kindly ask you to consider raising your score.

Reviewer_1RK1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes TNDMs, which model the mapping from training-data subsets to model performance with low-rank tensor networks (CP/TT), and computes Shapley/Banzhaf values and interactions from a single fit.

Critical strengths

Addressing non-additive effects among training data is meaningful, and the distinction between interaction degree and tensor rank is clearly presented.

Critical weaknesses

Limited novelty.

The paper applies standard tensor regression models to a new domain. In addition, there are already many papers that compute Shapley values for tensor product models ([1,2,3]). What advantages does the approach in this paper offer over these existing methods?

Expressivity of low rank.

Low rank may limit expressivity. Even an additive function has CP rank d, which exceeds the ranks used (≤16 for d = 24). Insufficient rank may also introduce spurious interactions.

Scale.

Real LLMs use massive data, but the only LLM retraining experiment uses 10 groups (40 documents). This is insufficient to demonstrate the method's effectiveness at realistic scale.

References.

[1]. Farzaneh Heidari, Chao Li, and Guillaume Rabusseau. Tractable Shapley Values and Interactions via Tensor Networks. In Proceedings of the 29th International Conference on Artificial Intelligence and Statistics (AISTATS), PMLR 300, pp. 3043–3051, 2026. arXiv:2510.22138.

[2]. Farzaneh Heidari and Guillaume Rabusseau. TN-SHAP-G: Graph-Structured Tensor Network Surrogates for Shapley Values and Interactions. In Proceedings of the 43rd International Conference on Machine Learning (ICML), PMLR 306, 2026. arXiv:2606.01540.

[3]. Seokhun Park, Insung Kong, Yongchan Choi, Chanmoo Park, and Yongdai Kim. Tensor Product Neural Networks for Functional ANOVA Model. In Proceedings of the 42nd International Conference on Machine Learning (ICML), PMLR 267, 2025. arXiv:2502.15215.

Questions for authors

Isn't the order of the training data also important, not just which training data is selected? If so, why was the training order not taken into account?

2 comments

We thank the reviewer. The revision clarifies when low-rank datamodels offer compact representations and adds evidence on attribution accuracy, spurious interactions, and natural-data interventions.

1. Joint removals and selection on natural data. On AG News classifier refitting, TT has the lowest two-group removal error among the methods in Table 25 at 256 and 512 observations: 0.144 and 0.094, versus 0.575 and 0.524 for the ProxySPEX expansion (approximately 4.0- and 5.6-fold reductions) and 0.424 and 0.290 for CP. The expansion is most accurate at 128 observations. Appendix K.2 gives the held-out removal protocol.

For RoBERTa–LoRA fine-tuning, TT has the lowest selection regret for subsets retaining three and five groups: 0.0131 and 0.0275, versus 0.0140 and 0.0294 for the strongest other method. RBF is best at size four and the ProxySPEX expansion at size six, and RBF or linear regression for removals. Table 26 reports all selection and removal sizes.

2. Higher-order terms in fine-tuning. On Qwen2-7B Sport–Music, CP reaches held-out R2=0.668R^2=0.668, compared with 0.565 for quadratic regression. Removing CP terms above degree two lowers its score to 0.634, so terms above degree two contribute 0.034 of this gain within the same fitted predictor (Section 4.3, Sport–Music paragraph, p. 8; Appendix K.2, Table 23).

3. Compositional dependencies. On the four-part relation task, TT reduces prediction NMSE from 0.251 to 0.0046 and joint-removal RMSE from 0.131 to 0.0061 relative to the ProxySPEX tree. Its Shapley, pairwise, and third-order errors are 0.071, 0.050, and 0.105, versus 0.175, 0.209, and 0.625 for the best method, with TT ahead in all three learner seeds. Appendix K.2 also reports pair-feature controls and the six- and eight-part tasks. On a separate eight-group retraining task, Section 4.4 and the revised Figure 4 (p. 9) show CP competitive with ProxySPEX at 64 observations and more accurate than the dedicated attribution estimators. At 128 observations, CP's three attribution errors are below 0.006, while the ProxySPEX expansion achieves the lowest errors.

4. Connection to theory and training regimes. Section 3 now explicitly distinguishes low rank from sparsity (final paragraph, p. 6). An expansion with ss nonzero coefficients in any product basis has CP rank at most ss. Our coverage construction instead has TT ranks at most two and CP rank 2m2^m, so every exact product-basis expansion requires at least 2m2^m coefficients (Appendix A.6, Theorems 5–6). This identifies dependencies that TT represents compactly despite requiring many sparse interaction terms. Section 4.4 adds an eight-hop transformer experiment to Section 4.2 (p. 9): at 128 observations, TT achieves NMSE 0.054 versus 0.788 and 0.799 for the ProxySPEX tree and expansion. Section 4.3 separately reports pretrained-model fine-tuning and classifier-refitting results.

5. Checkpoint dynamics. We agree that the complete-table analysis in Appendix J.15 and recovery from sampled outcomes are different questions. Appendix J.15 shows that, across checkpoints, interactions shift from synergy to redundancy while the non-additive part of the outcome table remains compressible; these interactions are computed from complete outcome tables. Appendix K.2 addresses the sampled setting in other tasks: TNDMs fitted to a fraction of retraining outcomes predict unseen subsets and joint removals, and on the four-part task they recover Shapley and interaction values more accurately than methods without pair features. Testing whether sampled fits at each checkpoint recover the same synergy-to-redundancy shift is future work.

Thank you for the helpful review. Changes are in blue in the revised PDF.

  • Novelty: Appendix A.6 adds a result specific to training data: complementary and interchangeable groups give TT ranks at most two but exponentially many coefficients in any product basis. The four-part relation learner has this setup, and TT lowers prediction/removal error from 0.251/0.131 (ProxySPEX tree) to 0.0046/0.0061 (Table 24).
  • Spurious interactions: on a retraining game with known zero interactions, TNDM-CP has the smallest spurious interactions of all compared methods (Appendix K.3, Table 27).
  • Training order: a batch-order test on a synthetic task shows that changing the order shifts held-out cross-entropy by only 0.0033 and 0.0021 on average (Appendix K.3).
  • Baselines: ProxySPEX, KernelSHAP, KernelSHAP-IQ and SVARM-IQ added (Tables 21–22).

We hope to have addressed your concerns and answered your questions. We remain available if you have any more questions during the end of the discussion period.

If we properly addressed your concerns and questions, we kindly ask you to consider raising your score.

Reviewer_E7raOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary
  • The paper looks beyond linear datamodels and studies interactions among subsets of training data.
  • It introduces Tensor Network Datamodels (TNDMs), which model the training outcome as a low-rank tensor function of the training subsets. Prior datamodels are linear, low-degree polynomial, or kernel-based.
  • The authors argue that interaction degree and tensor rank are distinct measures of complexity, and that tensor rank is the better-suited prior when training groups are complementary or redundant.
  • A single fitted TNDM is used to compute Shapley and Banzhaf values, higher-order interactions, and removal effects via tensor contraction, without additional retraining.
Critical strengths
  • Their theoretical framework for stylized reasoning-based data attribution establishes concrete links between tensor rank, interaction degree, and the structure of the training data. It gives a stylized explanation for why multi-hop and redundant data produce high-order but low-rank interactions.
  • The paper evaluates the approach across many setups, including reasoning tasks, fine-tuning, classifiers, and in-context learning. It reports metrics for different uses of a fitted datamodel, such as removal effects, interaction recovery, and data selection. The domain and metric coverage is satisfactory.
  • The method works well in experiments built to match its assumptions. On the controlled reasoning task, TNDMs outperform all baselines in prediction and joint-removal error, consistent with the theory.
Critical weaknesses
  • The methodological novelty is limited. TNDM is essentially TN-SHAP (Heidari, Li & Rabusseau, 2025) applied to data valuation instead of feature attribution. TN-SHAP represents a predictor as a factorized multilinear map over a coefficient tensor, which is nearly the same formulation as Definition 3. The difference is mainly the setting, not the machinery, and the related-work section discusses the overlap in only a few sentences.
  • The "fit one surrogate, then compute many attributions via contraction" pipeline was also introduced in TN-SHAP-G (Heidari & Rabusseau, 2026). TN-SHAP-G fits a surrogate from a small number of oracle queries and then recovers Shapley interaction indices through the multilinear extension without further model queries.
  • The core assumption, that the map from training subsets to outcomes is approximately low rank, is only weakly supported empirically. The main motivation seems to be the stylized reasoning model in Section 3. Experimental results are consistent with the assumption but are only indirect evidence for it. None of this evidence directly compares low rank against sparse alternatives.
  • The literature review and comparisons are missing related spare Möbius/Fourier work. This comparison matters for the paper's own examples. The authors should cite and position their approach against sparse-interaction methods:
    • Sparse Möbius Transform (Kang et al., 2024) and ProxySPEX (Butler et al., 2025). These works assume few nonzero interaction terms, whereas TNDM assumes structured but possibly dense ones.
    • Discrete signal processing on set functions (Püschel & Wendler, 2021) and learning set functions that are sparse in non-orthogonal Fourier bases (Wendler et al., 2021).
    • HarsanyiNet (e.g., Chen et al., 2023), Sparse Concepts in DNNs (Ren et al., 2023). Harsanyi dividends are exactly Möbius coefficients.
  • The functions in Theorem 1 are exactly mm-sparse in the Möbius basis, so a sparse model is as compact as CP rank mm. So it is not obvious that low rank beats sparsity.
  • Interaction baselines in the experiments are limited to generic regressors. For example, Fig. 4 evaluates pairwise and third-order Shapley interaction indices but does not compare against dedicated estimators such as KernelSHAP-IQ, SHAP-IQ, SVARM-IQ, Faith-Shap, or Shapley-Taylor.
  • The controlled experiments, which give the paper's strongest results, are designed to match the method's assumptions. For example, in the reasoning tasks, training groups are explicitly constructed to be complementary or redundant. This yields an exactly low-rank value function based on their theory. In other settings, the advantage of their method is smaller and less consistent.
Questions for authors
  • Could the authors provide direct evidence that low rank describes datamodels value functions better than sparsity? For example, they could report the singular-value spectra and the best rank-rr CP/TT approximation error. This approximation curve can be overlaid with top-KK Möbius and Fourier approximation curves at matched parameter counts. This analysis could be done easily on the value functions they have already fully enumerated.
  • Could the authors compare the Shapley values and interaction indices in Fig. 4 against dedicated Shapley-based estimators such as KernelSHAP, KernelSHAP-IQ, SHAP-IQ, and SVARM-IQ at matched budgets?
  • Could the authors also add a sparsity-based method like ProxySPEX as a baseline on at least the 24-group reasoning task and one fine-tuning task? ProxySPEX works with randomly sampled subsets, so it only requires fitting another surrogate to the already-collected retraining outcomes.
  • For tasks with many training groups dd, the experiments observe outcomes for only a small number of subsets, and the sampling design differs across experiments. Some experiments sample uniformly over all 2d2^d subsets, some impose cardinality constraints, and others use different designs. How were these designs chosen? Could the authors elaborate on how the main comparisons would change under a single design applied to all experiments?
  • Is there a sampling design with better sample complexity for the target functions? A kk-way AND is nonzero on only a 2−k2^{-k} fraction of subsets. Such basis functions are highly coherent, so uniform sampling may be inefficient for high-degree complementary supports. An ablation comparing sampling designs would also clarify how much of the reported estimation error comes from the design rather than the model class. For comparison, KernelSHAP samples from the Shapley kernel so that the regression objective matches the target quantity. Is it possible to consider a design matched to the target quantity?
2 comments

We thank the reviewer. The revision strengthens the distinction from sparse models and adds comparisons with ProxySPEX and dedicated attribution estimators.

1. Contribution and prior work. We agree that fitting Definition 3 is low-rank tensor completion. Section 2 gives a principled, unified view of datamodeling that makes degree bounds, sparsity, and tensor rank explicit, and Section 3 explains why low rank is a natural prior. Complementary and interchangeable training groups can generate high-order dependencies with compact tensor representations. TN-SHAP and TN-SHAP-G [1,2] explain a fixed model's predictions, while TNDMs predict outcomes after training on unseen subsets. Our contribution connects these training-data dependencies to tensor rank, while building on prior attribution computations. Section 5 now positions TNDMs against sparse Möbius/Fourier methods and Harsanyi dividends.

2. Low rank versus sparsity. We agree that Theorem 1 also admits a sparse Möbius representation. Appendix A.6 now formalizes the connection: an expansion with ss nonzero coefficients in any product basis has CP rank at most ss, since each basis element is separable. Consequently, high CP rank rules out sparse representations in every product basis, including Möbius and Walsh–Fourier. We then establish a separation: for a coverage function with mm required relations and two interchangeable providers per relation, keeping each pair adjacent gives TT ranks at most two, whereas the exact CP rank is 2m2^m. Every product-basis expansion over the 2m2m binary variables therefore requires at least 2m2^m nonzero coefficients.

For this analytical coverage function, TT uses 8m8m scalars; at m=12m=12, this is 96 scalars, while capturing 95% of centered variance requires 339,551 nonconstant Walsh terms (Appendix K.1). This extends the comparison beyond exact representation to approximation in the Walsh basis.

We also compare compression of measured learner outcomes. On the four-part learner, TT compresses the full outcome table to NMSE 0.0064, versus 0.096 for sparse Möbius and 0.182 for sparse Walsh—approximately 15- and 28-fold lower errors, using 56 TT core entries versus 48 sparse coefficients. Table 20 also reports the original tasks and affine residuals, where sparse methods have lower errors. Low rank thus captures a distinct structural regime rather than universally outperforming sparsity-based methods.

3. Prediction and joint removals. With 128 observations on the four-part learner, TT reduces prediction NMSE from 0.251 to 0.0046 and two-group removal RMSE from 0.131 to 0.0061 relative to the ProxySPEX tree, approximately 55- and 21-fold reductions (Appendix K.2, Table 24). The updated Table 1, now includes both ProxySPEX variants under the same observation and validation splits. At 1,024 observations, TT and the ProxySPEX expansion reach prediction NMSE of 0.071 and 0.072, respectively; the expansion is more accurate at smaller budgets.

4. Dedicated attribution baselines. The revised Figure 4 now includes KernelSHAP, KernelSHAP-IQ, SVARM-IQ, and both ProxySPEX variants. At 64 queries, CP's Shapley, pairwise, and third-order errors are 0.096/0.210/0.323, versus 0.209/0.584/0.840 for KernelSHAP-IQ—approximately 2.2-, 2.8-, and 2.6-fold reductions. CP is competitive with ProxySPEX at this budget; at 128 queries, CP's three errors are below 0.006, while the ProxySPEX expansion achieves the lowest errors. On the same game, CP's spurious pairwise and third-order interactions are 4.9- and 2.9-fold smaller than KernelSHAP-IQ's (Appendix K.3). Table 22 also includes KernelSHAP and SVARM-IQ. No single method is best across all settings. Our theory establishes dependencies with low TT rank that require exponentially many terms in any product-basis expansion (Appendix A.6). The four-part and eight-hop TT experiments illustrate the value of exploiting this structure in datamodeling (Appendix K).

5. Sampling strategies. A new paragraph in the Limitations notes that the tensor formulation enables sampling designs tailored to prediction, joint removals, and attribution, building on adaptive tensor completion [3]. Comparing uniform, subset-size-stratified, and adaptive designs is future work.

References

[1] Farzaneh Heidari, Chao Li, and Guillaume Rabusseau. Tractable Shapley Values and Interactions via Tensor Networks. AISTATS, PMLR 300:3043–3051, 2026.

[2] Farzaneh Heidari and Guillaume Rabusseau. TN-SHAP-G: Graph-Structured Tensor Network Surrogates for Shapley Values and Interactions. ICML, PMLR 306, 2026.

[3] Akshay Krishnamurthy and Aarti Singh. Low-Rank Matrix and Tensor Completion via Adaptive Sampling. NeurIPS 26:836–844, 2013.

Thank you for the helpful review. Changes are in blue in the revised PDF.

  • Low rank vs. sparsity: Appendix A.6 proves that when each of mm required relations can be learned from either of two groups, TT ranks are at most two, while any product-basis expansion (including Möbius and Walsh) needs at least 2m2^m coefficients.
  • Experiment matching A.6: the four-part relation learner has exactly this setup. TT compresses its outcome table with NMSE 0.0064 versus 0.096 (Möbius) and 0.182 (Walsh) (Table 20), and from 128 observations lowers prediction/removal error from 0.251/0.131 (ProxySPEX tree) to 0.0046/0.0061 (Table 24).
  • New baselines: ProxySPEX (Table 21), and KernelSHAP, KernelSHAP-IQ and SVARM-IQ (Fig. 4, Table 22). At 64 queries, CP's attribution errors are 2.2–2.8 times lower than KernelSHAP-IQ's.
  • Related work: Section 5 now covers sparse Möbius/Fourier methods and Harsanyi dividends.

We hope to have addressed your concerns and answered your questions. We remain available if you have any more questions during the end of the discussion period.

If we properly addressed your concerns and questions, we kindly ask you to consider raising your score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.