Executive Brief
Traditional AI often fails when 90% of data shows zero claims. This paper applies Tsetlin Machines to the French MTPL dataset, demonstrating how propositional logic captures "tail risk" more effectively than GLMs. Discover a framework that matches modern predictive power with the "audit-ready" transparency required by regulators.
What You’ll Read
Balance Benchmarks: Why 20–40% sampling is the "sweet spot" for stable, predictive rules.
Smart Interactions: Discovering complex risk couplings (e.g., Density × Vehicle Power) automatically.
100% Transparency: Attributing risk using Beeswarm plots without the noise of SHAP approximations.
Key Risk Proxies: Analyzing how population density drives accident and theft frequency.
