Inducing Human-Readable Logic Rules for Classification

243 concepts211 papers1 proposal

Modern deep learning models typically map inputs into high-dimensional continuous vector spaces, making their internal decision logic opaque. In contrast, human reasoning relies heavily on discrete, symbolic structures—conditional "if-then" statements, disjunctive normal forms (DNF), or relational rules. The core philosophy of interpretable rule induction is to automatically discover these symbolic structures directly from data or complex models. This bridges the gap between high predictive accuracy and human-verifiable reasoning.

As Ivan Bratko et al. observe in their retrospective on Inductive Logic Programming (ILP), ILP turns 20 - Biography and future challenges, representing learned models through formal logic provides an inherent transparency. Instead of interpreting a statistical approximation, domain experts can directly audit the specific causal or relational rules driving a prediction.

Direct Symbolic Optimization

To learn rules directly from raw data without relying on a neural network, algorithms must search the discrete space of logical combinations. Instead of performing gradient descent over continuous weights, these methods optimize the composition of logical operators (like AND and OR).

In Compositional noisy-logical learning, Alan Yuille et al. introduce a framework that structures binary classifiers as DNF expressions. By optimizing these logical compositions, their approach matches or outperforms dense statistical ensemble methods like AdaBoost while using up to 50 times fewer weak classifiers. This demonstrates that sparse symbolic logic can be just as expressive as complex continuous models.

Similarly, constraint learning focuses on inducing the underlying rules governing a system rather than merely predicting an output label. Luc De Raedt et al. show in Learning Constraints From Examples that expressing induced theories as logical rules allows models to complete partial data and reason about complex dependencies, offering a clear path for manual refinement by humans. To extend this capability to natural language understanding, Ole-Christoffer Granmo et al. propose the Relational Tsetlin Machine in A relational tsetlin machine with applications to natural language understanding. By learning first-order logic Horn clauses, the system builds compact, noise-robust knowledge bases that remain fully readable by human operators.

Post-Hoc Extraction and Global Surrogates

When a high-performing black-box model is already deployed, rule induction can serve as a distillation mechanism, translating dense mathematical operations into human-readable logic.

In high-stakes domains like clinical medicine, black-box decisions can be dangerous. To address this, Pietro Lio et al. developed PECLIDES Neuro: A Personalisable Clinical Decision Support System for Neurological Diseases, which extracts and optimizes a sparse set of logical rules from a trained Random Forest. The authors demonstrate that the collective intelligence of complex ensembles can be compressed into a few transparent queries suitable for clinical justification without sacrificing diagnostic accuracy.

For broader model maintenance, Owen Cornec et al. introduce AIMEE: Interactive model maintenance with rule-based surrogates. AIMEE extracts the global logic of an opaque model into Boolean decision rules in DNF. This global surrogate allows users to inspect the model's entire decision boundaries and directly modify its behavior by editing the rules.

On an individual prediction level, Sameer Singh proposes a local explanation method in Anchors: High-Precision Model-Agnostic Explanations. Rather than using linear feature-importance approximations, "Anchors" computes local if-then rules that define a precise coverage region. This tells the user exactly when the explanation holds true, significantly reducing the cognitive effort required to interpret the model's logic.

Neuro-Symbolic and Rule-Constrained Architectures

Rather than treating neural networks and symbolic rules as mutually exclusive alternatives, hybrid architectures embed logical constraints directly into deep learning frameworks.

In Deep Neural Networks Constrained by Decision Rules, Yuzuru Okajima et al. propose Rule-Constrained Networks (RCNs). This architecture replaces the standard continuous classification layer of a deep neural network with a selection mechanism that chooses from a predefined set of human-readable if-then rules. By forcing the network to output a rule rather than a raw probability vector, the system guarantees that every prediction is backed by an explicit logical justification.

Standard Network:  [Deep Features] ----> [Continuous Layer] ----> Opaque Label
Rule-Constrained:  [Deep Features] ----> [Selection Layer]  ----> Explicit Rule ----> Transparent Label

This neuro-symbolic integration is also highly effective for sequential data. Ruixuan Yan et al. introduce a framework for time series data in Neuro-symbolic Models for Interpretable Time Series Classification using Temporal Logic Description. By generating classification rules formatted as formal temporal logic, the system allows domain experts to verify, audit, and manually adjust parameters, ensuring that the model's data-driven insights align perfectly with established scientific knowledge.

History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_interpretable_rule_induction_0409d3ce,
  title = {interpretable rule induction},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/interpretable-rule-induction/rdr_0409d3cec0a54aee8d8c61130e540f68},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_interpretable_rule_induction_0409d3ce_v1,
  title = {interpretable rule induction},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/interpretable-rule-induction/rdr_0409d3cec0a54aee8d8c61130e540f68/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake