Split Conformal Prediction Intervals: Finite-Sample Coverage and Beta-Distributed Realised Coverage
Build a distribution-free prediction band from calibration scores and see why it covers the promised share of new points even when the model is wrong. Change the calibration size, the miscoverage level, the noise and the model, and compare the guaranteed coverage k over n plus one with the Beta distribution of realised coverage. A Blender-built plotting easel and rail gauge, with Monte Carlo, quadrature and quantile checks built in.
This simulator builds a tiny, real Factorization Machine — the pairwise-interaction mechanism behind production recommenders like Wide & Deep and DeepFM — and trains it live in front of you. Every categorical feature value (a user, an item, a device, a time of day) owns a 3D latent embedding you can watch move in the scene as genuine stochastic-gradient-descent updates pull related features together or push unrelated ones apart, while a linear "wide" term and a pairwise "FM" term combine into a single click-through-rate prediction you can inspect from either side.
Build a distribution-free prediction band from calibration scores and see why it covers the promised share of new points even when the model is wrong. Change the calibration size, the miscoverage level, the noise and the model, and compare the guaranteed coverage k over n plus one with the Beta distribution of realised coverage. A Blender-built plotting easel and rail gauge, with Monte Carlo, quadrature and quantile checks built in.
3D · Three.js / WebGL renderer · 60 FPS target · runs fully client-side, no install