Faithful Rule Learning for Tabular Data Cell Completion

ICLR 2026 Conference Submission14372 Authors

Keywords: Rule Learning, Tabular Data, Interpretability, Faithfulness, Datalog

TL;DR: We propose fully interpretable models for tabular data cell completion, accompanied by rule extraction algorithms. The extracted rules provide faithful explanations for the model's prediction.

Abstract: Tabular data cell completion aims to infer the correct constants that could fill a missing cell in a table row. While machine learning (ML) models have proven to be effective for this task, the limited interpretability restricts their applicability in trust-critical domains. In this paper, we develop two interpretable ML models to predict whether a candidate constant should fill the empty cell of an incomplete row by learning Datalog rules describing chain-like patterns of relations. Both models are *fully interpretable with formal guarantees*: we provide algorithms that take a model instance and extract an *equivalent* set of rules, in the sense that both the model and the rules produce the same output for any input table over a fixed relation schema. Furthermore, our models utilize different aggregation strategies to offer distinct trade-offs regarding expressive power and ease of rule extraction. Evaluations reveal that our models achieve state-of-the-art performance on tabular data cell completion with superior interpretability.

Primary Area: neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)

Submission Number: 14372