acceptodds
Withdrawn from ICLR 2027

Predict by Regions, Explain by Examples: Instance-Level Explanations for Tabular Prediction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

11 views

Abstract

Tabular prediction models are widely used in real-world decision-making, making instance-level explanations increasingly important alongside predictive performance. Existing explanation methods largely focus on feature attribution, while example-based explanations often retrieve nearby training instances without ensuring that those examples reflect the structure actually used for prediction. We propose TabERA (Tabular Explainable Region-based Architecture), a tabular prediction framework that grounds prediction and instance-level evidence in the same learned predictive regions. TabERA assigns each input to a region represented by a learned centroid and constructs its prediction from a shared regional component together with an instance-specific correction, yielding an explicit decomposition into a regional baseline and an individual adjustment. The same region is then used to retrieve observed training instances as evidence, while the retrieved cases remain outside the predictive computation. We evaluate TabERA on 21 OpenML classification datasets against 13 representative tabular prediction models. TabERA achieves competitive predictive performance while learning regions with clear label-relevant structure. Across datasets, the regional component accounts for a substantial portion of the predictive decision, and the learned regions generally contain sufficient observed instances to support region-grounded retrieval. These results show that instance-level examples can be tied directly to the model's predictive structure without making retrieved cases part of the prediction itself, providing a practical bridge between tabular prediction and example-based explanation.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Tabular Learning, Local Explanation, Example-Based Explanation, Instance Retrieval, Predictive Regions

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_e1yAOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TABERA, a tabular classification model that connects prediction and example-based explanation through a shared region structure. An encoder assigns each input to a learned centroid, and a linear head combines a regional baseline with an instance-specific correction. After training, examples are retrieved from the assigned region to contextualize the prediction, without participating in the predictive computation. Across 21 OpenML datasets and five splits, the paper reports competitive accuracy and F1 against 13 baselines, label-relevant regional structure, and faster prediction than TabR on 15 datasets. Two credit-classification examples illustrate the proposed explanation interface. The central question is whether this structural connection produces faithful explanations of the final prediction and offers an advantage over existing interpretable models.

Critical strengths
  1. Clear architectural connection between prediction and explanatory context. The regional baseline explicitly participates in prediction, and the same region assignment determines the candidate examples. The exact logit decomposition makes this connection inspectable. Separating prediction from example retrieval also allows evidence to be requested without requiring training-instance lookup for every prediction.

  2. Competitive reported predictive performance across multiple datasets. The evaluation covers 21 classification datasets, five splits, and a range of strong tree-based and neural baselines. The reported accuracy of 0.850 and F1 of 0.791 suggest that the regional architecture can retain useful predictive performance. The regional-label and retrieval-support analyses provide initial evidence that the learned groups are informative and usually contain enough examples for the proposed interface.

Critical weaknesses
  1. The experiments do not establish that the explanations faithfully reflect the final predictions.

    Sharing a region connects the retrieved examples to one component of the prediction, but does not establish that they explain the final decision. The displayed explanation does not explicitly account for the instance-specific correction. Section 6 acknowledges the absence of feature attribution and the dependence on region quality; nevertheless, the displayed regional characteristics have not been validated as evidence of what drives the model's output.

    Section 5.1 measures regional accuracy, class agreement, and retrieval coverage. The reported 3.96% class-change rate supports agreement between regional and final predictions on most samples, but does not measure changes in probabilities or margins. Figure 3(b) also shows regional accuracy of approximately 22% and final accuracy of approximately 84% on OpenML 1493. This dataset appears to account for nearly all of the reported 2.99-point average improvement from the correction, yet the paper does not explain how its regional evidence accounts for the final decisions in this case.

    The two examples in Section 5.3 illustrate the interface but do not validate its fidelity. No quantitative comparison with ordinary nearest neighbors or random same-region examples establishes the contribution of the proposed retrieval rule to explanation quality. The evidence therefore supports regional context, while the central contribution of explaining individual predictions remains insufficiently validated.

  2. The comparison with existing interpretable models is insufficient.

    TABERA trains its own classifier and provides an accompanying explanation mechanism. However, Section 4 compares predictive accuracy and F1, and Section 5.2 compares prediction speed with TabR; neither evaluates explanation quality against other interpretable models. Section 2.2 distinguishes TABERA from ProtoGate (Jiang et al., 2024), whose prototypes participate in prediction, but provides no experimental comparison. Interpretable tree, rule, and additive models are also absent from the explanation evaluation.

    The reported point estimates broadly support the authors' claim of competitive prediction: TABERA matches LightGBM in accuracy, is slightly below ModernNCA, and ties ModernNCA and ResNet in F1. This is a useful result, but without measurements of explanation fidelity, stability, or complexity, it leaves the comparative contribution of the explainable architecture unclear. The current evaluation does not establish whether TABERA offers better explanations at comparable predictive performance or a better accuracy-interpretability trade-off.

Questions for authors
  1. Where is the appendix referenced in the paper? Section 4.1 and the reproducibility statement refer to an appendix, but it is absent from the supplied PDF. Please provide the appendix and clarify whether it was submitted separately. Include the stated implementation details, hyperparameter ranges, and evaluation protocol.

  2. Why does the correction produce such a large accuracy gain on OpenML 1493? Figure 3(b) shows an approximately 62-point accuracy gain from the correction on this dataset, which appears to account for nearly all of the average gain. Please confirm the values, explain this behavior, and report regional accuracy, final accuracy, and class-change rates for each dataset. Please clarify how the explanation accounts for the final decisions on OpenML 1493.

  3. Where are TABERA's results in Table 1? Section 4.2 reports TABERA's accuracy of 0.850 and F1 of 0.791, but I could not find its experimental results in Table 1. Please provide the missing entry and the per-dataset results for TABERA and the baselines, including variability across the five splits.

  4. Do the explanations faithfully reflect the final predictions? Please provide quantitative evidence that the retrieved examples and displayed regional characteristics reflect the final prediction, including the instance-specific correction. Please compare explanation fidelity with ordinary nearest-neighbor retrieval and random same-region selection, and state the fidelity measure used.

  5. How does TABERA compare with existing interpretable models? Please compare TABERA with ProtoGate (Jiang et al., 2024) and representative interpretable tree or additive models on the same datasets. Report predictive performance and explanation quality using clearly defined measures, such as fidelity, stability, or complexity.

Reviewer_e8ZROpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes TabERA, a model for tabular prediction that splits data into regions using encodings. It is able to provide explanations using instances from the same region. It compares to 13 methods over 21 datasets.

Critical strengths

The paper provides a wide comparison over many methods and datasets.

Critical weaknesses
  • Table 1 (and subsequent results) report no confidence intervals, and I believe the TabERA results are missing entirely. Given the large number of datasets, this information would be important.
  • Missing ablations and details. Is the instance-specific correction justified? What are the trained βraw\beta_{raw} values?
Questions for authors

Clarification requests

  • I wonder about the justification of instance-specific correction. It is unclear why it is needed, whether βraw\beta_{raw} is shared or instance-specific, and what values it takes in the experiments. For dataset 1493 there is a big jump in performance after the correction, and I did not find an intuition for it or whether the model works worse when trained without it.
  • Why are tabular foundation models not included as baselines? They have good predictive power and are able to provide instance level explanations from attention.
  • What value of ρ\rho is used in the exponential averaging?
  • Does the model work worse when trained without the instance correction? Is Fig. 3b computed on models trained separately without the correction or just by removing a term from the inference?
  • When a region is too small (<8 samples) are outside training samples used for the explanation? Those samples do not influence the prediction, are they really useful as explanations?
  • How useful are the explanations to users/consumers? The samples in regions seem very diverse. It is unclear whether a user would find an explanation informative if it shows membership in a diverse region containing dissimilar individuals. The explanations may offer limited insight to end users; evidence of their usefulness would be valuable.
  • Revisit the claim that "explainability is an important requirement for deploying predictive models" (Rudin, 2019); I think this argument is the opposite of what Rudin wrote, focusing on interpretability instead.

Suggestions for future work

  • Discuss whether the framework can provide contrastive explanations, given the literature showing that people seek them (Miller, 2019)

Tim Miller, Explanation in artificial intelligence: Insights from the social sciences, Artificial Intelligence, 2019, https://doi.org/10.1016/j.artint.2018.07.007.

Changes that would raise my score

  • Please add confidence information for the results.
  • Address explainability, arguing for how it can be used.
  • Discuss privacy: whether training examples can be shown as explanations in highly regulated areas, and add this as a limitation.
  • I couldn't find the appendix that the paper refers to for details. Please add it.
Reviewer_2iBkOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes TabERA, a tabular classifier designed to provide example-based, instance-level explanations. Each input is assigned by cosine similarity to one of many centroids, and the logits decompose into a regional baseline and an instance-specific correction. After training, the nearest training instances within the assigned region are returned as evidence. They do not enter the prediction. On 21 OpenML classification datasets against 13 baselines, TabERA reports a competitive accuracy. The explanatory claims are supported by structural analyses, an inference-time comparison with TabR, and a qualitative case study on credit-g.

Critical strengths
  1. Competitive accuracy from a constrained, near-nearest-centroid architecture. Mean accuracy are on par with tuned gradient-boosted trees and ModernNCA. This holds even though predictions are largely determined by a region-shared term and prediction requires no access to stored training instances. The region bottleneck therefore costs little predictive performance.
Critical weaknesses
  1. The explanatory claim, which is the paper's main contribution, is not evaluated. The paper argues that region-grounded examples "reflect the structure actually used for prediction" better than post-hoc retrieval. However, no explanation-quality metric is reported (e.g., removal/retraining tests, neighbour-label agreement, simulatability) and no alternative example-selection method is compared: neither plain kNN in the same embedding, nor TabR/ModernNCA neighbours, nor prototype or representer-point methods, which the paper cites. Counterfactual and contrastive explanations, widely studied for credit and loan decisions such as the one used as motivation, are not discussed, and the retrieved evidence is one-sided. The only evidence is two hand-picked examples (Fig. 5). On credit-g, region-majority accuracy is only slightly above global-majority accuracy, so the highly pure regions shown are not representative of that dataset.

  2. The supporting analyses are largely implied by the architecture, and the empirical evidence lacks rigor. The paper has no theoretical contribution, so it rests on its experiments, which are thin. Region purity is compared only with the global-majority label, which any partition of a trained classifier's embedding would likely beat, and a k-means control on a plain MLP embedding is missing. Table 1 reports only means over five splits, with no standard deviations, ranks, or significance tests. The TabERA row is not visible in the rendered table. The appendix with implementation details and hyperparameter ranges, cited in Sec. 4.1 and the reproducibility statement, is absent.

Questions for authors
  1. Clarification: How do the region-grounded examples compare with plain kNN retrieval in the same embedding, and with TabR/ModernNCA neighbours, under an established explanation metric (e.g., neighbour-label agreement with the model's prediction, or prediction change after removing the retrieved instances and retraining)? A clear advantage would directly support the main claim and would raise my assessment.
  2. Clarification: What region-majority accuracy does k-means with ⌊N⌋\lfloor\sqrt{N}\rfloor clusters achieve on the penultimate embedding of a standard MLP trained on the same data? If it is close to 81.82%, the purity result in Fig. 3a does not distinguish TabERA's regions.
  3. Clarification: Please report per-dataset means and standard deviations for Table 1, and provide the missing appendix (encoder depth/width, hyperparameter ranges, whether baselines were rerun or taken from Lee et al., 2025).
  4. Clarification: Since rr is normalised in Eq. 3, the correction is independent of ∥q−ca∥\|q-c_a\|. How is this consistent with the "position within the region" narrative of Sec. 5.3, and how does the model behave on datasets with more classes than regions (e.g., 1493)?
  5. Suggestion: Retrieving contrastive cases from neighbouring regions with a different predicted label would provide counterfactual-style evidence and connect the method to the counterfactual-explanation literature.

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

Related papers

Discussion (0)

Sign in to comment.