From Black Box to Formula: KAN Networks for Auditable Commerce Scoring
Can a single-line mathematical formula replace a 64 MB neural network? We tested it experimentally — with surprising results.
Commerce scoring has a transparency problem. Neural networks deliver precise relevance scores, but nobody can explain why a particular product received a particular score. We tested whether a novel architecture — Kolmogorov-Arnold Networks (KAN) — can solve this. The Experiment Nexbid's current scoring combines two systems: a weighted formula (rule-based) and a fine-tuned DistilBERT (64 MB, ONNX, 23ms per request). Our question: Can KAN automatically extract the scoring formula from data — human-readable and verifiable? Unlike classical neural networks that learn fixed activation functions on nodes, KAN learns flexible functions on edges. These edge functions can then be converted into symbolic expressions: log, sqrt, x², and so on. Result: Formula Exactly Rediscovered We generated three synthetic datasets of increasing complexity (10,000 samples each, 8 features). The result for the linear scoring formula: Feature KAN Result Status Feature A (relevant) Weight correctly reconstructed ✅ Error < 0.005 Feature B (relevant) Weight correctly reconstructed ✅ Error < 0.003 Feature C (relevant) Weight correctly reconstructed ✅ Error < 0.005 Feature D (relevant) Weight correctly reconstructed ✅ Error < 0.004 Feature E (relevant) Weight correctly reconstructed ✅ Error < 0.005 Noise Feature 1 0 ✅ Correctly eliminated Noise Feature 2 0 ✅ Correctly eliminated Noise Feature 3 0 ✅ Correctly eliminated Note: Specific feature names and weights are part of our production scoring logic and are intentionally not published. This experiment validates the methodology, not the specific formula. 99.1% of learnable signal captured. Maximum coefficient error: 0.005. Three irrelevant features automatically eliminated. Training time: 12 seconds on CPU. What KAN Means for Commerce Three applications emerge: Auditability: Instead of "the score is 0.73," KAN says: score = 0.30 × relevance + 0.20 × quality +.... Publishers and advertisers can read the formula. Feature Discovery: KAN automatically reveals which signals actually matter — more valuable than the score itself. Infrastructure: A symbolic formula evaluates in <0.1ms vs. 23ms for DistilBERT. No ONNX runtime, no model hosting. The Limits For nonlinear formulas (with terms like x&sup0;·&sup8; or log(1+5x) ), KAN captures 97-98% of the signal as a learning model, but symbolic extraction loses accuracy. Feature interactions ( a × b ) are implicitly learned but not symbolically extracted as products. Bottom line: Perfect audit tool for linear formulas, strong feature importance tool for complex functions. The Vision: Computationally Verified AI Long-term, we are working toward a pipeline: Data → KAN → Symbolic Formula → Formal Verification …
- Author
- Holger von Ellerts
- Published
- 2026-03-19
- Topics
- Research, Open Source, Agentic