Home▸Data Science▸Factorization Machines 2D: Wide + Deep Feature Interactions

Factorization Machines 2D: Wide + Deep Feature Interactions

Interactive 2D Factorization Machine: pick a training example, train it live with real SGD on a logistic loss, and watch each feature's 2D latent embedding move on a pannable/zoomable plane while a wide (linear) bar chart and FM (pairwise) interaction lines combine into a CTR prediction.

Data Science2DAdvanced60 FPS📱 Mobile-adapted⇄ 3D version

Details

CategoryData Science

RenderingCanvas 2D

DifficultyAdvanced

Target FPS60

ControlsPick one value per field (User/Item/Device/Time) to form a training example, toggle the actual click label, switch the score view between Wide-only, FM-only or combined, adjust the learning-rate and auto-train-speed sliders, run a live SGD training step (or toggle continuous auto-train), randomize all weights, and drag/scroll the embedding plane to pan and zoom.

factorization machineswide and deepdeepfmrecommender systemsgradient descentctr prediction

Related simulations

Nesterov Momentum: Contour Map & Convergence RaceA 2D contour-map view of plain gradient descent, classical momentum, and Nesterov…

Data Science: A Physics PerspectiveExplore the core concepts of data science – analytics, modeling, and machine learning –…

Nesterov Momentum: Look-Ahead Gradient on an Ill-Conditioned ValleyRace plain gradient descent, classical momentum, and Nesterov accelerated gradient down an… ⊞ All 144 in Data Science →

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 2D latent embedding you can watch move on a pannable, zoomable plane as genuine stochastic-gradient-descent updates pull related features together or push unrelated ones apart, while a linear "wide" bar chart and a pairwise "FM" interaction-line view combine into a single click-through-rate prediction, tracked live on a scrolling loss sparkline.

⚙ Under the hood

Train a real Factorization Machine live in 2D: pick a user/item/device/time example, label it clicked or not, run genuine SGD updates, and watch each feature's 2D latent embedding move on a pannable/zoomable plane while a wide (linear) bar chart, FM (pairwise) interaction lines, and a live loss sparkline combine into a click-through-rate prediction.

factorization machineswide and deepdeepfmrecommender systemsgradient descentctr prediction

2D · HTML5 Canvas 2D · 60 FPS target · runs fully client-side, no install