Concept
Recommendation System
The algorithms that decide what content, products, or connections to surface to each user - personalising feeds, search results, and product listings based on inferred individual preferences, past behaviour, and the behaviour of similar users.
Added May 18, 2026
Recommendation systems are among the most economically impactful AI applications. Netflix estimates that its recommendation system generates $1 billion of value annually by reducing churn. Amazon attributes 35% of its revenue to its recommendation engine. TikTok's For You Page, YouTube's recommended videos, Spotify's Discover Weekly, and LinkedIn's People You May Know are all recommendation systems that collectively shape much of what people watch, buy, listen to, and connect with.
Collaborative filtering is the foundational approach, based on the observation that users who have agreed in the past tend to agree in the future. User-based collaborative filtering finds users similar to you (similar rating patterns) and recommends items they liked that you have not seen. Item-based collaborative filtering finds items similar to items you liked (similar co-rating patterns) and recommends them. Matrix factorisation decomposes the user-item interaction matrix into latent user and item embedding vectors whose dot products predict ratings, efficiently compressing patterns in massive interaction matrices.
Content-based filtering recommends items similar to items the user has previously engaged with, based on the item's own features: a movie recommendation system might recommend other sci-fi films to a user who watches sci-fi, using genre, director, cast, and plot features. Content-based methods can recommend new items without cold start problems (they do not need any user interaction with the item) but are limited by the features available to describe items.
Deep learning has largely replaced classical matrix factorisation in industrial systems. Neural Collaborative Filtering (NCF) uses MLPs to learn complex non-linear user-item interaction functions. Wide & Deep (Google) combines a linear model (memorising specific feature interactions) with a deep network (generalising to new feature combinations). DeepFM and DCN (Deep & Cross Network) automatically learn feature interactions without manual feature engineering.
Sequential recommendation models the user's evolving preferences over time: rather than treating all past interactions equally, they weight recent interactions more heavily and model the sequence of interactions to infer current intent. Session-based recommendation handles short-term sessions without long-term user history. Transformer-based sequential models (BERT4Rec, SASRec) achieve state-of-the-art results by modelling complex temporal patterns.
Graph-based recommendation represents users and items as a bipartite graph (users connected to items they have interacted with) and uses GNNs to propagate preference information through the graph. Pinterest's PinSage, Alibaba's NGCF, and Uber Eats' recommendation GNN are industrial deployments of this approach.
The exploration-exploitation tradeoff (the multi-armed bandit problem) arises directly in recommendation: exploit known preferences to maximise immediate engagement, or explore new content that might better serve the user. Systems that only exploit become filter bubbles; systems that explore excessively frustrate users.
Analogy
A deeply knowledgeable librarian who has watched every reader who has ever visited the library, knows what they have checked out, and has noticed that readers who loved book A almost always love book B, that readers who loved C and D tend to love E, and that readers new to the library who seem like Alice tend to want what Alice reads. This librarian gives each visitor personalised recommendations that synthesise patterns across the entire reading community - scaled to millions of users and billions of items by recommendation algorithms.
Real-world example
Spotify's Discover Weekly generates a personalised playlist of 30 songs every Monday for each of its 400 million users. The system uses collaborative filtering (users who listen to similar artists as you also liked these songs), audio content analysis (deep learning embeddings from the audio signal capturing tempo, mood, energy), natural language processing on playlist names and music blogs (capturing genre and style descriptions), and sequential modelling of each user's listening history. Discover Weekly has become one of Spotify's most-loved features, with billions of streams generated by its recommendations.
Why it matters
Recommendation systems are how most people experience AI in daily life - these algorithms determine the content of social feeds, the products seen in online shopping, the songs on playlists, and the shows on streaming platforms. Their design choices have enormous societal implications for what gets promoted, filter bubbles, and engagement addiction. Understanding how they work - collaborative filtering, deep learning architectures, the exploration-exploitation tradeoff - is essential for anyone building consumer applications or thinking critically about AI's influence on information consumption.
In the news
No recent coverage - check back later.
Related concepts