Logic Based Learning via Continuous Feature Binarization

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 2164 papers187 conceptsfrozen 2026-10-04

Most real-world data is continuous: a patient’s temperature, a stock’s price, or the expression level of a gene. However, human reasoning and formal logic are fundamentally discrete. We don't typically reason about a temperature of 38.62°C; we reason about the presence of a "fever."

A growing body of research focuses on bridging this gap by transforming continuous numerical features into Boolean "atoms." This process, known as binarization or discretization, allows researchers to apply the rigor of symbolic logic and the efficiency of bitwise operations to complex, high-dimensional datasets.

The Intuition of Thresholding

The core mechanism of this research direction is the conversion of a real-valued feature into one or more binary indicators. The simplest approach is thresholding: if a value exceeds a certain point, the bit is "1"; otherwise, it is "0."

While simple, the choice of threshold is critical. Dmitry M. Malioutov and Kush R. Varshney demonstrate in Exact Rule Learning via Boolean Compressed Sensing that quantile-based discretization—splitting data into equal-sized bins—can transform continuous features into a truth table. This allows the search for interpretable classification rules to be reframed as a signal processing problem, providing theoretical guarantees that the resulting rules are the "simplest" possible explanations for the data.

In clinical settings, these thresholds are often driven by biological "normal" ranges. Sergio Daga and colleagues in the GEN-COVID Multicenter Study used binarization to standardize heterogeneous patient data. By converting granular lab results into binary indicators of organ system damage, they were able to use hierarchical clustering to identify five distinct COVID-19 phenotypes that were previously obscured by the noise of continuous measurements.

Discovering Optimal Boundaries

Manually selecting thresholds is not always feasible. Recent work focuses on data-driven binarization, where the model itself learns the most informative split points.

Zihao Chen and Xiaomeng Wang propose using the internal nodes of decision trees as feature extractors in An Interpretable Loan Credit Evaluation Method Based on Rule Representation Learner. By treating every split condition in a tree as a binary attribute, they transform tabular data into a structured set of indicators that capture the most statistically significant boundaries. This approach allows case-based reasoning systems—which usually require discrete inputs—to scale to modern datasets without manual knowledge engineering, as noted by Francesca Toni and Guilherme Paulino-Passos in their technical report on case relevance.

Why Binarize? Interpretability and Efficiency

The move toward binary representations is driven by two primary goals: making AI understandable to humans and making it faster for machines.

1. Human-Readable Logic Black-box models like deep neural networks are powerful but opaque. By binarizing data, researchers can use algorithms that produce "Or’s of And’s" (Disjunctive Normal Form). Cynthia Rudin and Tong Wang argue in Learning Optimized Or's of And's that these logical expressions are small enough for a human to memorize, making them ideal for high-stakes environments like medicine or law. Similarly, Litao Qiao and Weijia Wang show in Rethinking Logic Minimization for Tabular Machine Learning that binarized tabular data can be processed by logic minimization techniques to achieve accuracies comparable to XGBoost while remaining entirely transparent.

2. Computational Speed Once data is binarized, learning becomes a matter of bit manipulation. This is the foundation of the Tsetlin Machine, a logic-based learner. Ole-Christoffer Granmo and Jivitesh Sharma have demonstrated that by converting continuous inputs into bit-strings, Tsetlin Machines can outperform traditional neural networks in precision and speed for tasks ranging from forecasting disease outbreaks to solving contextual bandit problems.

Furthermore, binarization enables the use of SAT solvers—highly optimized tools for checking Boolean satisfiability. Florent Avellaneda leverages this in Efficient Inference of Optimal Decision Trees, showing that transforming numerical data into a binary feature space allows SAT solvers to find "optimal" decision trees (those with minimum depth) orders of magnitude faster than previous methods.

Robustness Through Simplification

Counter-intuitively, throwing away the "precision" of floating-point numbers often makes models more robust. In genomic studies, Andres R. Masegosa found that binarizing gene expression into "high" and "low" categories filters out experimental noise and outliers, leading to more stable predictions of genetic interactions.

This "denoising" effect is a recurring theme. Whether it is Dongsheng Yuan using binarization to create platform-agnostic cancer classifiers or Akram Aldroubi using binary similarity matrices for subspace clustering, the evidence suggests that the presence or absence of a signal is often more critical for defining identity than its exact magnitude. By reducing the resolution of the input, these methods reveal the underlying logical structure of the problem.

Papers in this version

Showing 60 of 164 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 186 papers
  2. Version 1 · 2026-10-04 · 164 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_feature_discretization_2de33169,
  title = {feature discretization},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/feature-discretization/rdr_2de3316907134edba4063e2d6e16f906},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_logic_based_learning_via_continuous_feat_2de33169_v1,
  title = {Logic Based Learning via Continuous Feature Binarization},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/logic-based-learning-via-continuous-feature-binarization/rdr_2de3316907134edba4063e2d6e16f906/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}