ECSEL: Explainable Classification via Signomial Equation Learning

Adia Lumadjeng

Modern machine learning often trades explainability for predictive performance. Deep neural networks achieve remarkable accuracy but obscure the reasoning behind predictions. In high-stakes domains such as medicine, finance, and human resources, understanding the underlying decision process is as important as accuracy. Surprisingly, simple, well-chosen formulations can capture the essential structure in data, delivering both accurate predictions and transparent, human-readable explanations. We present ECSEL, an explainable classification method that learns closed-form equations in the form of signomials. The key idea is to unify prediction and explanation: instead of relying on post-hoc interpretability, ECSEL directly produces a mathematical expression that both classifies and explains the data. We show that many problems admit compact signomial structure, which ECSEL can efficiently recover. On symbolic regression benchmarks, ECSEL identifies a larger fraction of ground-truth equations than state-of-the-art methods while requiring substantially less computation. Leveraging this efficiency, it achieves competitive classification performance with standard machine learning models without sacrificing interpretability. Beyond predictive performance, ECSEL provides analytical insights into model behavior, including global feature effects, decision-boundary structure, and local attributions. Experiments on benchmark datasets and real-world case studies in e-commerce and fraud detection demonstrate that the learned equations reveal dataset biases, support counterfactual reasoning, and offer actionable insights.

Short Bio

Adia Lumadjeng is a third-year PhD student at the University of Amsterdam. She holds an MSc in Econometrics (Operations Research and Logistics tracks) from Erasmus University Rotterdam, as well as an MSc in Science Education and Communication and a BSc in Industrial and Applied Mathematics from TU Delft. Her research focuses on the intersection of operations research and machine learning, with a particular emphasis on explainable AI for fraud detection. She also has an interest in mathematical physics. Her work explores how structured mathematical models can be combined with modern learning techniques to produce interpretable and efficient solutions. The work presented here is recent and currently under review. She lives in Amsterdam together with her dog Ellie from İzmir, Turkey.

Venue

Friday, April 3rd, 2026, 4:00 pm

Online

The link for the seminar is: https://teams.live.com/meet/9391563036995?p=tdbRXvyfTxcbvkVFj3