AI & Data › Machine Learning Basics

Recommendation System

Suggesting items based on behavior and similarity.

Also known as: recommendation system, recommender, recommendations

A recommendation system predicts which items a user is likely to engage with and orders candidates accordingly: products, videos, articles, connections. It works from signals about users and items — past interactions, attributes, and similarity between either — and typically runs as a pipeline that narrows a large catalogue to a short, ranked list.

catalogue (millions) → candidate retrieval (hundreds) → ranking model → top-N shown

Two families of signal drive most systems. Collaborative approaches infer relevance from what similar users did; content-based approaches match item attributes to a user’s history. Real systems blend them, and most of the engineering effort goes into the candidate and ranking stages, data freshness and evaluation.

The classic mistakes:

• Optimising offline accuracy alone. A model that predicts past clicks well can still produce a dull or repetitive experience. Measure the outcome users and the business care about, online.
• Feedback loops. Showing only what the model already favours starves the system of data about everything else. Reserve exploration on purpose.
• Popularity bias. Popular items dominate every list, and new or niche items never get a chance. Account for it in both ranking and evaluation.
• Ignoring cold start. New users and new items have no history. Plan a fallback based on attributes, context or editorial choices.
• Evaluating on random splits. Recommendations depend on time; split by time to avoid training on the future.

Where to start: a simple popularity or recency baseline, then a candidate generator, then a ranker, with an online experiment to confirm that offline gains are real.