Modern recommendation engines represent every user and every item (a movie, a product, a song) as a point — a vector — in a shared, high-dimensional embedding space. Items that people tend to like together end up close together in that space. A user's own position is built from the items they have interacted with, and their collaborative-filtering recommendations are simply the nearest items to that point.
This scene compresses that idea down to 3 dimensions you can fly around. Each coloured cluster of dots is a genre neighbourhood in the catalog. The glowing indigo marker is a simulated user. As you feed it interaction history, watch it drift out of the "cold start" centre of the space and toward the neighbourhood it has been engaging with — pulling the recommended items (bright, connected by beams) along with it.
Real systems like Netflix's and Spotify's use embeddings with hundreds of dimensions, but the geometric idea is identical to this 3D scene: recommend whatever is nearest to you in the learned space, then deliberately mix in a little distance so the model keeps discovering new signal instead of only repeating itself.
An interactive 3D user-item embedding space where adjusting a simulated user's interaction history shows collaborative-filtering recommendations shifting toward nearby items in the embedding space.
Users and items live as points in a shared learned space; recommendations are simply the nearest items to a user's position, and an explore/exploit slider trades relevance for discovery of new tastes.
Pick a genre the simulated user has been watching, raise history strength to move them out of the cold-start centre, then tune Top-K and exploration to see the recommendation beams update live.
Production recommenders use hundreds of embedding dimensions, but the underlying geometry is exactly this: nearest neighbours in a learned space, with a pinch of randomness for exploration.