Recommender system

Recommender systems rank or select items for particular users, rather than merely retrieving every item that matches a query. They are used in e-commerce, streaming media, online publishing, advertising, education, and social platforms. The recommendation may be a ranked list, a single suggestion, or an explanation of why an item was selected.¹

Inputs include explicit feedback, such as star ratings or likes, and implicit feedback, such as purchases, dwell time, skips, searches, or repeated plays. Because implicit signals are plentiful but ambiguous, a click may indicate interest, curiosity, or accidental selection. The system therefore estimates preference rather than observing it directly.

Recommendations can reduce information overload and help users discover obscure or new items. They also influence visibility: an item that is recommended more often can receive more interactions, creating feedback loops that amplify existing popularity or earlier predictions.

Collaborative filtering recommends items from patterns shared among users or items. User-based methods find people with similar behavior, while item-based methods identify items that tend to receive similar responses. Matrix factorization represents users and items with latent numerical factors, a method that became influential through the Netflix Prize era.²

Content-based filtering uses item attributes—such as genre, ingredients, subject terms, or text—to recommend items resembling those a user previously engaged with. Hybrid systems combine collaborative and content signals, often adding context such as device, location, time, or current session. Modern systems may use deep neural networks, sequence models, or transformer architectures, but a more complex model is not automatically more useful or fair.

Many large services separate candidate generation from ranking. A fast first stage narrows millions of items to a manageable set; a second stage applies a richer model and business or safety constraints. Re-ranking can promote diversity, freshness, novelty, or editorial priorities alongside predicted relevance.

Offline evaluation commonly compares predicted rankings with held-out interactions using measures such as precision, recall, mean reciprocal rank, or normalized discounted cumulative gain. These measures are convenient but incomplete: logged behavior reflects earlier recommendations, and a model can score well while narrowing choice or failing under changing conditions.

Online experiments, including A/B tests, measure outcomes such as useful engagement, retention, purchases, or satisfaction. Responsible evaluation also examines calibration, coverage, diversity, serendipity, and performance across demographic or activity groups. User studies can reveal harms that click-through rates miss.

Bias may enter through unequal historical exposure, popularity concentration, missing data, or feedback loops. Privacy concerns arise when systems infer sensitive characteristics from apparently ordinary behavior. Governance measures include data minimization, access controls, audit logs, documentation, user controls, explanations, and ways to contest or correct consequential recommendations. Research frameworks distinguish personalization from manipulation and call for evaluation beyond immediate engagement.³

Recommender systems do not always optimize for an individual prediction; some are designed to support a conversation. Conversational recommenders ask questions, incorporate stated constraints, and revise suggestions after feedback, making preference elicitation part of the task. Group recommenders instead seek a compromise for several people, as in shared entertainment or travel planning.

The cold-start problem has several forms. A new user supplies little behavior, a new item has no interaction history, and a changing domain can make old preferences unreliable. Exploration strategies address this uncertainty by deliberately showing some less-certain items, connecting recommendation to the multi-armed bandit problem and reinforcement learning.⁴

Long-tail recommendation is another difficult edge case: niche items may be highly relevant to a small audience but receive too little data to rank well. Diversity-aware methods, editorial signals, knowledge graphs, and human feedback can help. In sensitive domains such as health or news, a technically accurate ranking may still require stronger provenance, safety review, and limits on personalization.

Collaborative filtering

A method that infers preferences from patterns of interaction among users and items.

Content-based filtering

A method that recommends items whose attributes resemble those of items previously preferred.

Cold-start problem

The difficulty of making useful recommendations for new users, new items, or rapidly changing environments.

Implicit feedback

Behavioral evidence such as clicks, purchases, views, skips, or listening time that indirectly signals preference.

Matrix factorization

A family of methods that represents users and items through lower-dimensional latent factors.