A recommendation system is an information filtering infrastructure that predicts and surfaces items, content, or actions likely to be of interest to a specific user, based on behavioural history, explicit preferences, item features, or combinations thereof. It encompasses collaborative filtering approaches that exploit user-item interaction patterns, content-based methods that match item attributes to user profiles, and hybrid models that combine multiple signals. Modern recommendation systems employ deep learning, graph neural networks, and large language models to capture complex preference patterns at scale. They are commercially critical infrastructure in e-commerce, streaming media, social networks, and digital advertising.
Content
- The discipline emerged in the mid-1990s with collaborative filtering systems such as Tapestry and GroupLens, which applied the intuition that users with similar past behaviour have similar future preferences. Amazon’s item-to-item collaborative filtering patent (2001) established scalable real-time recommendations as commercially valuable infrastructure, and Netflix’s $1M Prize (2006-2009) catalysed algorithmic advances in matrix factorisation.
- Three primary paradigms dominate production systems. Collaborative filtering decomposes the user-item interaction matrix into latent factor vectors and finds approximate nearest neighbours. Content-based filtering builds user profiles from item attribute vectors and matches them at inference time. Hybrid systems — prevalent in industrial deployments — combine both with contextual signals (time, device, location) and session-aware sequence modelling using recurrent or transformer architectures.
- At scale, recommendation systems face engineering challenges distinct from modelling: candidate retrieval from billion-item catalogues using approximate nearest-neighbour indices; multi-stage ranking pipelines that progressively score and re-rank thousands of candidates; feature engineering pipelines that join real-time events with precomputed embeddings; and A/B testing infrastructure to evaluate policy changes safely against live traffic.
- In 2024-2025, large language models are being integrated as zero-shot and few-shot rankers, enabling semantic understanding of item descriptions and user intent without extensive interaction data. Graph neural networks processing user-item interaction graphs are demonstrating strong performance on cold-start and long-tail recommendation. Conversational recommendation, where users express preferences through dialogue, is emerging as a paradigm that blends language models with traditional ranking, particularly in voice-assistant and chatbot contexts.