Every product a retailer sells can be represented as a point in an abstract embedding space — a set of learned coordinates where similar items sit close together. A shopper is placed in that same space based on their browsing and purchase history. The engine's job is simple to state and hard to do well: find the nearest points to the shopper and show those first.
This lab visualises two ways of building that space. A content-based layout clusters products purely by category — electronics near electronics, books near books. A collaborative filtering layout instead clusters items that real customers tend to buy together, regardless of category — a coffee machine might sit near a specific mug or a cookbook because people who bought one also bought the other.
Large marketplaces report that recommendation modules can drive a large share of page views and a meaningful share of revenue; blending collaborative and content-based signals (a "hybrid" recommender) generally outperforms either approach alone, especially for new or low-history shoppers — the classic "cold start" problem.
A 3D map of a product embedding space where a shopper's position and a blend of collaborative-filtering and content-based signals determine which items get recommended.
Products cluster two ways: by category (content-based) or by real co-purchase patterns (collaborative filtering). Blending the two changes which items sit nearest the shopper — and therefore which get recommended.
Pick a shopper persona, then adjust the collaborative-filtering weight, the number of recommendations (k), and diversity. Watch the highlighted, connected products change along with predicted AOV and conversion lift.
Hybrid recommenders that mix collaborative and content-based signals typically outperform either approach alone, and are especially important for new shoppers with little history — the "cold start" problem.