acceptodds
Under review as a conference paper at ICLR 2027

Faithful Rule Learning for Tabular Cell Completion

Abstract

Tabular cell completion aims to infer values that can fill missing cells in incomplete table rows. While machine learning models have achieved strong performance on this task, their limited interpretability restricts their applicability in high-stakes domains. In this paper, we introduce two interpretable models for tabular cell completion based on path-based relational patterns. Given an incomplete row and a candidate value, the models aggregate evidence from paths connecting the known constants in the row to the candidate value. A key feature of our approach is that the learned models admit *faithful* symbolic characterizations in the form of equivalent Datalog programs: for every model instance, we show how to extract an equivalent Datalog program that produces exactly the same outputs on every database over the underlying schema. We further analyze the relationship between the two models, showing how different aggregation mechanisms lead to distinct trade-offs between expressive power and simplicity of rule extraction. Experimental results demonstrate strong performance on tabular cell completion benchmarks while additionally providing human-readable explanations with formal faithfulness guarantees.

Primary area:
interpretability and explainable AI
Keywords:
Rule Learning, Datalog, Tabular Data, Faithfulness

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_5asFOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes MC and MC-max for missing-cell completion in relational tables. Both models use paths from known values to possible completions. MC adds the path contributions. MC-max uses the largest path contribution. The paper gives Datalog rules with the same outputs as each model. It also gives a separate method for explaining MC outputs on a fixed database. The analysis covers expressive power and rule extraction cost. The tests use tabular data and binary relations. The test data include values not seen during training.

Critical strengths

The extracted global rules give the same completions as the models. The guarantee holds on every database at matching prediction and extraction thresholds. The paper also gives a separate extraction method for MC on a fixed database. Its completeness guarantee applies only to that database.

The proposed models achieve the highest reported mean accuracy and PR-AUC on each tabular dataset. The best settings vary by dataset and metric. The models do not lead on every metric. MC-max also extracts rules quickly in the reported tests.

Critical weaknesses

Both models build on earlier path-based rule learning. They extend this approach to higher-arity tables through pairs of column values. I find the added modeling contribution over prior faithful rule learning unclear.

The tests compare the models with hypergraph methods and LLMs. They omit closely related faithful rule-learning methods. The gains over earlier interpretable methods remain unclear.

Questions for authors

I would appreciate the authors' responses to the following points.

  1. The extension gives faithful rules for higher-arity tables. Its paths still use pairs of column values. The added modeling power over rule learning on binary projections with column labels is not clear to me.

  2. The binary tests compare only MC and MC-max. How do these models compare with earlier faithful rule learners in prediction quality and rule complexity?

  3. The models pool evidence across observed columns with shared transition weights. Does this limit their use of source column roles in the tested tasks?

Reviewer_sv46OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

This paper studies faithful rule learning for tabular cell completion. It proposes two path-based models, MC and MC-max, that predict missing values by aggregating relational evidence across tables. The main contribution is a rule extraction framework with formal guarantees. The extracted Datalog rules produce exactly the same predictions as the learned models. Experiments on three tabular benchmarks and inductive knowledge graph datasets show competitive predictive performance.

Critical strengths
  • The paper establishes exact equivalence between the learned models and the extracted Datalog rules. This is a meaningful theoretical property that goes beyond approximate explanations.

  • The proposed methods achieve competitive accuracy and precision against several hypergraph baselines. The reported extraction times, particularly for MC-max, also suggest that the approach is computationally feasible on the evaluated datasets.

Critical weaknesses

W1. Limited evidence of a substantial new contribution.

Faithful rule extraction has already been studied [1, 2], including max-based variants [2]. The main extension here appears to be higher-arity relations. However, Eq. (1) represents these relations through pairwise adjacency matrices, which can lose joint tuple information. It is unclear how much genuinely higher-order reasoning the proposed models can capture beyond aggregating binary relational paths. The paper does not provide sufficient evidence to establish the significance of this extension over existing faithful rule-learning methods.

W2. Insufficient empirical validation.

The proposed models do not consistently outperform HC-MPNN in F1. For example, the best reported F1 on WP-IND is 68.8, compared with 73.2 for HC-MPNN, although the proposed methods achieve better accuracy and PR-AUC. More importantly, the evaluation primarily relies on balanced positive/negative candidate verification rather than completion over the full candidate space. The binary-relation experiments also lack direct comparisons with closely related methods, such as DRUM [3] and its faithful variants, including MMDRUM [2].

Furthermore, the reported extraction times for MC concern database-specific rule extraction rather than the full universally faithful extraction established in Theorem 1. The paper also does not demonstrate whether the extracted rules improve practical data auditing or completion. Thus, the practical significance of the formal guarantees beyond theoretical equivalence remains insufficiently established.

[1] Faithful Approaches to Rule Learning. KR 2022.

[2] Faithful Rule Extraction for Differentiable Rule Learning Models. ICLR 2024.

[3] DRUM: End-To-End Differentiable Rule Mining On Knowledge Graphs. NeurIPS 2019.

Questions for authors
  1. Since the models rely on pairwise projections of higher-arity relations, what dependencies can they capture beyond aggregating binary relational paths? A controlled example would help clarify the actual expressive power and limitations of the approach.

  2. The proposed models have limited F1 advantages over HC-MPNN. Can the authors provide stronger empirical evidence of their practical benefits, for example, through evaluation over the full candidate space or by demonstrating how the extracted rules support data auditing?

  3. How do the proposed methods compare with existing faithful rule-learning models, such as DRUM/MMDRUM, particularly on the binary-relation benchmarks? What additional capabilities does the higher-arity extension provide?

  4. For MC, the reported extraction times appear to concern database-specific rules. Can the authors clarify the practical feasibility of extracting the full universally faithful rule set guaranteed by Theorem 1?

1 comment

Thank you for the detailed response, the new example in Section 3.1, and the extra comparison in Table 11. These updates help clarify how the model works. However, I still have strong reservations about the overall contribution. I agree with Reviewer 5asF's assessment that the modeling advance over prior faithful rule-learning methods remains quite incremental.

Breaking high-arity tables down into pairwise column connections loses joint tuple information. Even in Table 2, most of the learned rules are still simple binary patterns. Summing paths from multiple cells in a row is a natural step, yet it feels like a modest extension compared to the heavy formalism used to present it.

There is also a clear gap between the theory and practical utility. Your response confirms that extracting the universal rule set for MC is too expensive to run in practice. To get around this, you fall back on database-specific rules, which lose global interpretability. Alternatively, MC-max allows fast extraction, but it drops the multi-cell summation that makes the model different from single-source methods in the first place.

Finally, the empirical validation remains incomplete. The evaluation still relies only on balanced candidate verification rather than full-space completion. In Table 11, tuning the rank for your models while comparing against baselines with a fixed rank also makes the comparison less direct, and the claimed benefits for data auditing remain untested.

While the theoretical proofs are sound, the practical and modeling gains are too limited for acceptance. I will therefore maintain my original score of 2.

Reviewer_huAwOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies table cell completion. The paper claims to provide two interpretable models based on path-based relational patterns. The models take an incomplete row and a candidate value as input. The models can be translated to Datalog language and provides exactly the same results. The results do not show significant improvement over baselines.

Critical strengths

The problem studied is clearly stated with Figure 1.

Critical weaknesses
  1. The problem studied is not new, and I do not see the novelty of the method.
  2. The method is not explainable as claimed. As far as I understand, section 4 states that the method can be translated into Datalog language, and the logical solver will give the same results as the model. But this is not surprising and nothing to do with the explainability of the model itself.
  3. The author try to present with loads of equations and theorems. But the method behind is simple. Trying to write a complex story only makes the paper hard to read.
  4. The experiments are not convincing. The proposed method is not superior compared with baselines. And the experiments do not show the explainability.
Questions for authors

See above.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.