A recommendation system, or recommender system, is software that predicts a user's preferences and suggests items, such as products, movies, music, or news articles, that the user is likely to want. Recommendation systems power much of the modern internet's personalization, from e-commerce product listings to video and music streaming feeds and social media timelines, and are among the most economically significant applications of machine learning.
Techniques
Two classical approaches dominate the field. Collaborative filtering infers a user's preferences from the behavior of similar users, without needing to understand item content directly. Content-based filtering instead recommends items similar in attributes to what a user has liked before, using features of the items themselves. Netflix's 2006-2009 open competition to improve its collaborative-filtering accuracy by 10%, with a $1 million prize, was a landmark event that popularized matrix-factorization techniques and drew mainstream attention to the field; it was won in 2009 by a team combining multiple models through ensembling. Modern systems increasingly use deep learning, representing users and items as vectors, or embeddings, learned so that similar users and items sit close together in a shared space, which can then be searched efficiently using a vector database and techniques related to semantic search. Neural network architectures such as two-tower models and, since the 2020s, transformer-based sequence models that treat a user's history as a sequence to predict the next likely interaction, have become common at large platforms.
Applications and business impact
Recommendation systems are central to the business models of Amazon, whose product recommendations drive a large share of sales, Netflix and YouTube, whose content recommendations drive most viewing, Spotify's music discovery, and social platforms such as TikTok and Instagram, where feed-ranking algorithms are widely credited with driving user engagement and advertising revenue.
Criticism
Recommendation systems have drawn criticism on several fronts. Feed-ranking algorithms optimized purely for engagement have been linked to filter bubbles and polarization, since maximizing watch time or clicks can favor sensational or extreme content. Critics have also raised concerns about algorithmic bias, where systems trained on historical interaction data can reinforce existing disparities in what gets recommended to whom, and about the opacity of these systems, an issue studied under explainable AI and interpretability research, since users and regulators often cannot see why a particular item was recommended. Privacy concerns are also common, since effective personalization typically requires collecting detailed behavioral data. As large language models have matured, some platforms have begun exploring conversational, LLM-based recommendation interfaces that let users describe what they want in natural language rather than relying solely on implicit behavioral signals.