A new arXiv preprint extends the Ex-Fuzzy Python library to regression for the first time, producing models of 10 to 15 human-readable rules that hit a mean R² of 0.86 across 10 benchmark datasets S¹. Every rule reads like a sentence a person could write. The question is whether that kind of transparency can survive contact with real, messy, regulated data, or whether 0.86 on clean academic benchmarks tells you anything useful about the day a hospital or a bank actually tries to deploy it.

My read: Rule-based AI is pushing back toward the mainstream. What I like here is the compactness: 10 to 15 rules is small enough that a human can actually read and audit them, which is the whole point. What I don't buy yet is the generalisability. The R² of 0.86 is self-reported across 10 KEEL datasets, which are clean academic benchmarks. The baselines (linear regression, MLP, random forests) may not be optimally tuned, so the comparison tells you more about the fuzzy method's own viability than about whether it beats alternatives. Until someone runs this on a real clinical or financial dataset with properly tuned baselines, the promise is real but unproven.

Why regression changes everything for fuzzy rules

Fuzzy logic has been around for decades. The idea is simple: instead of a hard threshold where "temperature above 40 is high," you let categories blur at the edges, so 39 is mostly high and 41 is entirely high. A fuzzy rule base is a set of if-then statements using those blurred categories, and a human can read every one of them.

The Ex-Fuzzy library, created in January 2024 and hosted on GitHub under an AGPL licence, has been building tools for this kind of explainable AI in Python P³. Until now, it handled classification: sorting things into categories. Regression, predicting a continuous number, was missing S¹.

That gap matters. A model that classifies a loan application as "approve" or "deny" is useful, but a model that predicts the expected default amount, or a patient's expected recovery time in days, is often what you actually need. And in regulated settings, you need to explain the number, not just produce it.

How the method actually works

The paper's core contribution is a way to build fuzzy rules that predict numbers instead of categories. It uses Mamdani fuzzy inference, a classic framework where each rule's output is a fuzzy set that gets combined and defuzzified into a single prediction S¹. The extension lets those output sets be scalar values learned directly from the training data.

The trick is in how the input space gets divided. The authors use Fuzzy C-Means clustering, an algorithm that groups data points into overlapping clusters rather than hard buckets, to find natural divisions in the data S¹. But they do it in what they call a target-aware way: instead of clustering only on the input features, they augment the space with the output variable, so the clusters form around regions that matter for prediction S¹. The result is linguistic variables (the "high," "moderate," "low" labels) shaped by what actually drives the output, not by the distribution of inputs.

They tested two shapes for those fuzzy sets: Gaussian (the classic bell curve) and trapezoidal (flat-topped with sloping sides). Gaussian consistently won S¹.

What the benchmarks show, and what they don't

The authors evaluated their method on 10 regression datasets from the KEEL repository, a standard collection of academic benchmarks S¹. They compared Gaussian and trapezoidal partition strategies against three baselines: linear regression, multilayer perceptron, and random forests S¹.

The headline number: Gaussian partitions achieved a mean coefficient of determination (R²) of approximately 0.86 S¹. R² measures how much of the variance in the target variable your model captures, with 1.0 being perfect. An R² of 0.86 across 10 diverse datasets is solid for an interpretable model.

But the paper does not claim to outperform the black-box baselines in accuracy S¹. The comparison is there to show the fuzzy method is competitive, not dominant. And the baselines may not reflect optimally tuned implementations, a risk the preprint itself does not fully address.

The rule bases came out compact: 10 to 15 rules per model S¹. That is the number a person could sit down and read in a few minutes, which is the entire value proposition.

What to do about it

If you work in a regulated domain where model explanations are mandatory, this is worth watching. Consider a small credit union that needs to predict expected loss amounts for loan portfolios. A random forest might give you a number, but when a regulator asks why a particular prediction came out at $12,000, you are stuck with feature importance scores that are approximate and hard to defend. A 12-rule fuzzy model that says "if credit score is low and loan-to-value is high, expected loss is $11,800" gives you something you can put in a report and sign your name to.

The Ex-Fuzzy library is open source under AGPL v3.0 and available on GitHub now P³. The regression extension described in the preprint is not yet merged into the main library as of this writing, so you would need to check the repository for the experimental code. If you want to test the concept, pick a small regression problem you already understand well, build a fuzzy model, and compare the rules against what your domain knowledge tells you should be driving the output. The rules either make sense or they don't, and that immediate check is something no neural network gives you.

What we don't know yet

This is a preprint, not a peer-reviewed paper S¹. Every performance number is self-reported by the authors and has not been independently verified. The R² of 0.86 is an average across specific KEEL datasets and may not hold on data with different noise profiles, dimensionality, or non-linear structure.

The paper discusses safety-critical applications as motivation but does not demonstrate deployment or validation in any real-world regulated setting S¹. The baselines (linear regression, MLP, random forests) may not be optimally tuned, which means the competitive comparison is suggestive rather than definitive.

The Ex-Fuzzy library had 57 stars on GitHub at the time of writing P³, which means the community is small and the code has not seen the kind of broad stress-testing that larger projects benefit from.

The next signal: watch the Ex-Fuzzy GitHub repository for a merged regression module and any issues filed by external users testing it on real data. If the authors submit to a journal like IEEE Transactions on Fuzzy Systems, the peer-review process will pressure-test the baseline comparisons. Subscribe and we will check both claims against whatever comes next.

Sources: S1 — Interpretable Fuzzy Rule-Based Regression Extension for Ex-Fuzzy Libra · P2 — Interpretable Fuzzy Rule-Based Regression Extension for Ex-Fuzzy Libra · P3 — Fuminides/ex-fuzzy · P4 — CLAUDE.md at main · Fuminides/ex-fuzzy · P5 — sageteamorg/scikit-fuzzy-c-means

More from Not A Tech Guy

Generated from an audited evidence pack with primary-source research. Social-media items are discussion signals, not verified facts. Nothing here is financial, legal or medical advice.