acceptodds
Under review as a conference paper at ICLR 2027

CART: Contextual Cell Reliability for Robust Tabular Prediction

Abstract

Real-world tabular data is rarely uniformly reliable: individual entries (cells) can be corrupted or mis-recorded while the rest of a record remains trustworthy, yet most robust tabular learning methods act at the sample level (train the whole model to be robust), the feature level (weight entire columns), or attempt to reconstruct every cell indiscriminately. We argue that none of these directly answers the question a predictor implicitly has to resolve at inference time: *how much should this particular observed cell be trusted for this particular record?* We propose CART, an architecture that (i) estimates a contextual, per-cell reliability score from full-record attention, (ii) reconstructs a true leave-one-out (LOO) counterfactual representation of every cell using an attention query that never observes the cell itself, and (iii) adaptively fuses the observed and reconstructed representations according to the learned reliability before classification. Across eight tabular benchmarks, three seeds, and six corruption severities, CART attains the best noise-robustness AUC on 7 of 8 datasets and the best or tied-best accuracy at 20% corruption on 6 of 8 datasets, while remaining competitive on clean data. A seven-way mechanistic ablation shows that the improvement is not explained by indiscriminate reconstruction, a fixed repair ratio, a random gate, or even a gate with the correct value distribution but permuted cell assignment (two-sided Wilcoxon signed-rank test across 8 datasets, in all three comparisons): the gain requires assigning the *correct* reliability score to the *correct* cell. Compared with reconstruction-oriented baselines built on the same tokenizer (a denoising autoencoder, a transformer autoencoder without LOO, and a variational autoencoder), CART attains the best cell-level reliability detection on roughly half of the datasets and is a close competitor elsewhere, showing that generic reconstruction quality and cell-reliability detection are related but distinct capabilities. Finally, CART's robustness advantage persists, though it does not always dominate, under corruption magnitudes and feature shifts not seen during training.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
tabular learning, robust machine learning, cell reliability, cell corruption, contextual reconstruction, robust prediction

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_9yQmOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses cell-wise corruption in tabular data, where individual entries may be erroneous while other entries in the same record remain reliable. The authors propose CART, a framework for robust tabular classification that combines contextual cell-level trust estimation, leave-one-out (LOO) reconstruction, and reliability-guided fusion of observed and reconstructed representations. Training uses synthetic corruption masks, reconstruction supervision, and a teacher–student consistency objective. Experiments on eight datasets report generally improved robustness relative to prediction-oriented baselines, while reconstruction diagnostics and inference-time interventions examine the proposed mechanism.

Critical strengths
  • Cell-level corruption is a meaningful setting, and distinguishing whether a value should be trusted from whether it can be reconstructed provides a useful motivation.
  • The trust, repair, and prediction branches have understandable roles. As formulated in Eq. (2), the repair branch excludes the target cell’s observed value from both its query and its attention inputs, preventing direct copying.
Critical weaknesses
  • The central mechanistic claims exceed what the ablations establish. The seven-way ablation in Section 5.4 modifies the fusion mechanism of an already trained CART checkpoint. Since the classifier has adapted to the original fusion distribution, performance degradation under raw inputs, reconstruction-only inputs, or substituted gates may reflect representation mismatch and component co-adaptation. These interventions establish dependence on the trained configuration, but do not establish its superiority over independently trained alternatives. Moreover, TabAE-E2E is presented as the control isolating LOO, yet Tables 2 and 5 report only its detection and reconstruction performance. Its downstream classification accuracy and NR-AUC are missing. The authors should retrain the key variants, particularly the raw-input and fixed-gate models, using their respective fusion configurations throughout training.
  • The necessity of contextual and cell-specific reasoning is not demonstrated. Access to full-record attention does not establish that contextual information drives the improvement. The synthetic corruption protocol may also create detectable marginal distribution changes. Relevant controls include a context-free cell detector, a feature-level gate, and corrupted-context interventions. The permutation procedure also requires clarification: permutations across different features can disrupt feature-specific fusion preferences. Within-feature permutations across records would more directly test record-specific assignment. Additionally, the similar corruption rates across SHAP-importance groups in Section 6 are expected under independent random corruption and provide limited evidence for the claimed conceptual distinction. The authors should train context-free and feature-level gating variants under matched training conditions.
  • Section 5.6 reports a mean Acc@20% of 79.5% for default CART, whereas the eight values in Table 1 average to 79.815%, approximately 79.8%, consistent with Table 3. This discrepancy affects the stated contribution of the training schedule and requires clarification. The loss weights, precise consistency objective, teacher construction, and data-splitting procedures are also insufficiently specified. The reported Wilcoxon tests concern checkpoint interventions; they do not establish statistical significance against the independently trained baselines in the main prediction table.
  • Some related works are not discussed. For example, RVAE [1] explicitly models cell-level outlier probabilities and repairs mixed-type tabular data. Outliers [2] distinguishes good outliers, i.e. novel data, from bad ones, i.e. those introduced by errors. These precedents do not invalidate CART’s potential contribution, but they require a sharper explanation and evaluation of what LOO repair and task-conditioned fusion add. RFOD [3] reformulates anomaly detection as a series of feature-wise conditional reconstruction tasks.
    • [1] Eduardo S, Nazábal A, Williams C K I, et al. Robust variational autoencoders for outlier detection and repair of mixed-type data[C]//International Conference on Artificial Intelligence and Statistics. PMLR, 2020: 4056-4066.
    • [2] Chen S, Fan W, Jin R. Outliers: The Good, the Bad and the Ugly[J]. Proceedings of the ACM on Management of Data, 2025, 3(4): 1-30.
    • [3] Ang Y, Yao P, Bao Y, et al. RFOD: Random Forest-Based Outlier Detection for Mixed-Type Tabular Data[C]//2026 IEEE 42nd International Conference on Data Engineering. IEEE, 2026: 2192-2205.
Questions for authors
  • Under matched training conditions, how do independently retrained raw-input, fixed-gate, and no-LOO (TabAE-E2E) variants compare with CART on Acc@20% and NR-AUC?
  • Can context-free and feature-level gating controls, together with within-feature trust permutations across records, establish the contributions of contextual reasoning and cell-specific assignment?
Reviewer_a4FsOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces CART, a method for robust classification of noisy records based on noise-injected cells (Gaussian noise draws, category swaps). The authors' work overlaps with prior un-cited work, and they introduce a transformer, LOO-repair similar to DDC, and a trust region. The authors benchmark CART against a few methods on traditional data science datasets not directly suited for the task of determining cell trustworthiness. They then ablate a portion of inference-time model-specific parts of the CART mechanism and show that the CART system as a whole is better than the components that were ablated.

Critical strengths
  1. It is admirable that the authors have published and shared their code and results on open science, which aids in verifiability of the paper's claims.
Critical weaknesses
  1. The same generator is used for the training noise that is used for the testing noise. The greatest problem with the paper and its methodology is that the "trust" head is exposed to exactly what it is meant to identify. The authors point out the problem of circularity (Introduction section, "A natural concern is circularity: if reliability decides how much to trust a cell, and the model is also trained to be robust, is it doing anything beyond curve-fitting the training corruption?") and propose a solution, but their solution does not address the problem. This work reduces to an extensive architecture built around a Gaussian denoiser. The problem is indeed circular, the paper identifies this, and attempts to make it seem as though the methodology addresses the problem. It does not.
  2. Real-world error does not often reduce to Gaussian noise. Problems like fat-fingering, typos, and systematic errors like joins and drops in data handling are more common and can often be correlated. Randomly perturbing an independent subset of the data is a good start but does not represent a meaningful use case. The authors state in the Introduction that the work is motivated by a specific example: rare mistyped income in patient data. The paper never addresses this use case.
Questions for authors
  1. You inject noise in a controlled manner that does not represent a real-world scenario. How does CART perform when tested on peer-reviewed benchmarks like Hospital, Beers, Tax, etc from Raha and Baran and CleanML for prediction?
  2. What happens to CART's performance when you swap the noise signature? I.e. - how does CART perform when you train on Gaussian-distributed contamination but test on data that has been contaminated with heavy tailed noise? Data with sentinel values? Data with scale errors? Other kinds of contamination?
  3. Why is Alqallaf et al. 2009 from "Annals of Statistics" not cited when your methodology uses the "Independent Contamination Model" from their paper?
  4. While "Detecting Deviating Data Cells" by Rousseeuw and Van den Bossche 2018 does not explicitly mention LOO, their method is effectively CART's core mechanism without the neural network. Their work and mechanism (DDC, that goes un-cited in your manuscript) masks the target cell in question as well as the other deviating attributes for a specific row. You added 7 mechanistic ablations but not a systematic ablation. Given this, what does the neural machinery actually add to CART? Does the transformer add anything over a DDC baseline?
  5. You supervise the trust head with the contamination mask, and contamination for test data uses the same generator that the train data uses. How do you know CART is learning reliability and not simply the output distribution of the noise generator?
Reviewer_iKu1OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes CART, a neural architecture for tabular classification that estimates the reliability of individual cells and reconstructs potentially corrupted values using the remaining features. The authors evaluate the method on only 8 datasets and report better robustness under synthetic corruption compared to several baselines.

Critical strengths
  • The problem of tabular data is relevant and practical
  • The paper is easy to follow
Critical weaknesses
  • Novelty -- The main ideas, such as synthetic feature corruption, corruption detection, and feature reconstruction, have already been explored in self-supervised learning for tabular data.
  • The authors do not provide enough discussion of previous work on tabular self-supervised learning, e.g., VIME, SCARF, SAINT, and others.
  • Missing baselines. Important self-supervised learning methods, such as VIME, SCARF, and SAINT, are not included in the experimental comparisons.
  • Limited evaluation The experiments are limited to 8 binary classification datasets. No regression tasks are considered. Consider of using an established benchamrks protocols like tabarena
  • All experiments use synthetic corruption. It is unclear how well the proposed method generalizes to real-world data errors and different types of corruption
Questions for authors

See the weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.