📊 statistics

ECSEL: Explainable Classification via Signomial Equation Learning

The paper introduces ECSEL, an efficient and interpretable classification method that learns compact signomial equations to achieve competitive accuracy while providing global and local explanations, bias detection, and counterfactual reasoning capabilities.

Original authors: Adia Lumadjeng, Ilker Birbil, Erman Acar

Published 2026-05-27

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Adia Lumadjeng, Ilker Birbil, Erman Acar

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a computer to make decisions, like spotting a fake credit card transaction or guessing if a website visitor will buy something. Usually, we use "black box" models (like deep neural networks). These are like brilliant but silent magicians: they get the answer right, but if you ask, "How did you know that?" they just shrug. You get the result, but no explanation.

This paper introduces ECSEL (Explainable Classification via Signomial Equation Learning), a new method that acts more like a transparent accountant than a silent magician. It doesn't just give you a "Yes" or "No"; it hands you the actual math equation it used to make that decision.

Here is how it works, broken down into simple concepts:

1. The Secret Ingredient: "Signomials"

Most AI models try to fit data into complex, tangled webs. ECSEL, however, looks for a specific, simpler pattern called a signomial.

Think of a signomial like a recipe for a smoothie.

Instead of a messy list of ingredients, a signomial recipe looks like this:
- (Strawberries raised to the power of 1.2) × (Bananas raised to the power of 0.5) × (Milk raised to the power of -0.3)
In this recipe, the numbers (exponents) tell you exactly how much each ingredient matters.
- If the number is positive, adding more of that ingredient makes the smoothie "stronger" (increases the score).
- If the number is negative, adding more makes it "weaker."
- If the number is big, that ingredient is the star of the show.

ECSEL learns these recipes automatically. Because the math is so clean, you can read the equation and instantly understand the logic: "Oh, high transaction amounts combined with low account balances are the main reason this looks like fraud."

2. Why It's Better Than the "Black Box"

The paper claims that while other AI models are often more accurate but impossible to understand, ECSEL is a happy medium.

The "Magic" vs. The "Map": A black-box model is like being driven to a destination by a driver who won't tell you the route. ECSEL is like giving you a GPS map that shows the exact turns.
Speed and Efficiency: The authors tested ECSEL on standard "puzzle" datasets (where the computer has to guess the hidden math formula). ECSEL solved these puzzles much faster and more often than the current state-of-the-art AI methods. It found the right "recipe" 96% of the time, while other methods only found it about 59% of the time, and ECSEL did it in a fraction of the time.

3. What Can You Do With the Equation?

Because ECSEL gives you the actual formula, it offers three superpowers that black boxes don't have:

The "What-If" Machine (Counterfactuals): You can ask, "What if this customer had spent $10 more?" Because the math is explicit, ECSEL can calculate the new answer instantly without retraining the whole model. It's like adjusting the ingredients in your smoothie recipe and immediately knowing the new taste.
The "Why" Button (Feature Importance): It tells you exactly which factors pushed the decision one way or the other. In a fraud case, it might say, "This transaction was flagged because the amount was huge (positive effect) but the account balance was zero (negative effect)."
Global vs. Local: It can explain the general rules for everyone (Global) and the specific reasons for one person's decision (Local) using the same equation.

4. Real-World Tests

The authors tested ECSEL in two real-world scenarios to see if it holds up:

Online Shopping: They tried to predict if a website visitor would buy something. ECSEL found a clear rule: "If the user looks at high-value pages but doesn't leave quickly, they are likely to buy." It was just as good at predicting the sale as complex AI models, but it gave a clear, readable reason why.
Fraud Detection: They tested it on millions of fake mobile money transactions. ECSEL learned a formula that spotted fraud by looking at specific patterns (like large transfers from accounts with zero balance). It was slightly less "accurate" at catching every single fraud case than a giant tree-based model (XGBoost), but it was much more precise (fewer false alarms) and, crucially, it provided a human-readable explanation of why it flagged a transaction.

The Bottom Line

ECSEL is a tool that refuses to trade clarity for performance. It proves that you don't need a mysterious, unexplainable AI to get good results. By sticking to a specific, mathematically elegant type of equation (the signomial), it builds models that are fast to train, competitive in accuracy, and fully transparent—like handing you the blueprint instead of just the finished house.

Technical Summary: ECSEL – Explainable Classification via Signomial Equation Learning

Problem Statement

Modern machine learning often prioritizes predictive performance at the expense of explainability. Deep neural networks achieve high accuracy but function as "black boxes," obscuring the reasoning behind predictions. In high-stakes domains such as medicine, finance, and human resources, understanding the underlying decision process is as critical as accuracy. While Symbolic Regression (SR) offers inherently interpretable models by generating explicit mathematical formulas, conventional SR approaches (e.g., Genetic Programming) rely on computationally expensive stochastic heuristics that struggle with high-dimensional datasets and fail to exploit prevalent structural patterns in data. Furthermore, existing SR methods are often designed as general-purpose tools rather than specialized classifiers, and post-hoc explanation techniques (e.g., LIME, SHAP) applied to black-box models can be unreliable or misleading.

Methodology

The paper introduces ECSEL (Explainable Classification via Signomial Equation Learning), a method that learns formal expressions in the form of signomial equations to serve as both a classifier and an explanation.

Core Formulation

A signomial function is defined as a finite sum of power-law terms with real-valued coefficients and exponents:
$z(x) = \sum_{k=1}^{K} \alpha_k \prod_{j=1}^{m} x_j^{\beta_{k,j}}$
where $x$ is the feature vector, $\alpha_k$ are coefficients, and $\beta_{k,j}$ are exponents.

ECSEL Architecture:

Class-Specific Scores: For a multi-class problem with $C$ classes, ECSEL learns a specific signomial score function $z_c(x)$ for each class $c$ .
Probability Transformation: Class probabilities are derived by applying a softmax function over the class-specific signomial scores. In binary cases, a sigmoid is applied.
Preprocessing: Since real-valued exponents require positive arguments, features are preprocessed via affine transformations (e.g., scaling to $[1, 10]$ ) to ensure $x_{ij} > 0$ .
Training Objective: The model is trained by minimizing cross-entropy loss with an $\ell_1$ regularization term applied to the exponents ( $\beta$ ). This encourages sparsity, driving irrelevant exponents to zero and resulting in compact, interpretable equations with fewer active features.
Optimization: The authors employ gradient-based optimization (Adam with gradient clipping) combined with numerical stability techniques (logarithmic transformations). For symbolic regression tasks, a multi-start staged optimization strategy is used (Adam for structure discovery, followed by L-BFGS refinement).

Theoretical Foundations

The paper establishes that signomial functions possess the Universal Approximation Property on compact subsets of the positive orthant. By mapping signomials to exponentials of linear functions via logarithmic transformation, the authors prove that signomials can approximate any continuous function to arbitrary precision, placing them theoretically alongside neural networks as universal approximators but tailored to multiplicative power-law relationships.

Interpretability Properties

ECSEL derives three levels of explanation directly from its closed-form structure, avoiding the need for sampling-based post-hoc methods:

Global Feature Behavior:
- Elasticity (G1): The log-log derivative of the class score with respect to a feature provides a closed-form measure of proportional sensitivity. For single-term models ( $K=1$ ), this is constant and equal to the exponent.
- Counterfactual Reasoning (G2): The effect of scaling a feature by a factor $q$ can be computed exactly without retraining, using the formula $z_{new} = \sum q^{\beta} z_{old}$ .
- Gradient-Based Sensitivity (G3): First-order responses to small proportional changes are derived analytically.
Decision-Level Effects:
- Decision Boundary Sensitivity (D1): The influence of features on the margin between classes is characterized by the difference in log-gradients.
- Probability Competition (D2): The sensitivity of predicted probabilities to feature changes is derived exactly, showing how features compete via normalization.
Local Feature Attribution:
- Log-Space Additivity (L1): For single-term models, the log-score is exactly additive in log-space, analogous to SHAP values.
- Gradient-Based Local Attributions (L2): For multi-term models, local attributions are obtained via first-order linearization around a baseline, providing exact closed-form gradients rather than estimated approximations.

Key Contributions

Identification of Signomial Structure: The authors identify signomial functions as a prevalent structure in symbolic regression benchmarks (e.g., 45% of AI Feynman equations) and propose ECSEL to exploit this structure directly.
Efficient Recovery: ECSEL recovers target equations in SR benchmarks more frequently than state-of-the-art approaches (DGSR, NGGP, NeSymRes) while requiring substantially less computation time.
Competitive Classification: ECSEL achieves classification accuracy competitive with established machine learning models (Logistic Regression, Random Forest, XGBoost, SVM, MLP) without sacrificing interpretability.
Formal Interpretability: The method satisfies analytical properties enabling global feature behavior analysis, decision-boundary analysis, and local feature attributions, all derived from model parameters rather than sampling.
Real-World Utility: Through case studies in e-commerce and fraud detection, the authors demonstrate that ECSEL's learned equations expose dataset biases, support counterfactual reasoning, and yield actionable insights.

Experimental Results

Symbolic Regression Benchmarks

Evaluated on the AI Feynman, Livermore, Jin, and Korns datasets (58 equations total):

Recovery Rate: ECSEL achieved a global average recovery rate of 95.86%, significantly outperforming DGSR (59.10%), NGGP (58.54%), and NeSymRes (56%).
Computation Time: ECSEL averaged 86.4 seconds per equation, compared to 612.9s (DGSR), 468.7s (NGGP), and 126.3s (NeSymRes).
Numerical Accuracy: Even when exact symbolic recovery failed, ECSEL typically yielded near-perfect numerical approximations ( $R^2 \approx 1$ ).

Classification Benchmarks

Evaluated on 11 standard binary and multi-class datasets (e.g., ILPD, COMPAS, LOAN, SKINNONSKIN):

Performance: ECSEL achieved the highest F1-score on 4 of 11 datasets and ranked within one percentage point of the best method on 9 of 11 datasets.
Minority Class Recall: ECSEL demonstrated superior performance in handling class imbalance, achieving significantly higher minority recall on datasets like ILPD (42.42% vs. 6.06% for XGBoost) and TRANSFUSION (41.67% vs. 38.89%).
Efficiency: While training times were higher than some baselines (e.g., Logistic Regression), the cost is shifted from explanation to training; post-training, all attributions are obtained in closed form at negligible cost, unlike black-box models requiring expensive SHAP/LIME computations.

Case Studies

Online Shopping Intention: ECSEL learned a 7-feature signomial identifying "PageValue per ExitRate" as the dominant predictor. It achieved competitive minority recall (75.7%) and provided exact counterfactual reasoning and local decompositions, outperforming LIME in stability and computational cost.
PaySim Fraud Detection: On a dataset with 6.3 million transactions (0.13% fraud rate), ECSEL achieved an F1 score of 79.08% with 94.27% precision. The learned equation revealed a "superlinear wealth-targeting mechanism" (exponent 1.42 on origin account balance), indicating fraudsters disproportionately target high-value accounts. ECSEL's performance was competitive with Deep Symbolic Classification (DSC) while adhering to stricter methodological validity (no temporal data leakage).

Significance and Claims

The paper claims that ECSEL bridges the gap between the interpretability of Symbolic Regression and the practical demands of modern classification. By combining computational efficiency with transparent decision-making, ECSEL offers practitioners an actionable alternative to black-box models in high-stakes domains. The authors emphasize that the method does not merely approximate black-box behavior but provides inherently interpretable models where the decision logic is explicit, mathematically rigorous, and capable of supporting counterfactual reasoning and bias auditing. The work suggests that simple, well-chosen formulations can capture essential data structures, delivering both accurate predictions and human-readable explanations.