Side Information
A recommender built only on clicks has nothing whatsoever to say about an item nobody has clicked yet. Side information is everything you know about users and items that does not come from the interaction matrix, and it is the usual way out of that hole.
What counts as side information
- Item attributes: genre, brand, price, the text description, the cover image, an audio embedding.
- User attributes: declared preferences at signup, locale, device, coarse demographics.
- Relational structure: a knowledge graph linking films to directors, products to categories.
- Context: time of day, what else is already in the basket.
How models absorb it
The defining property is availability for an entity with zero interactions — precisely when the collaborative signal is absent. A factorization machine folds attributes in as additional feature slots, learning an embedding per feature so feature interactions are modelled exactly as user-item interactions are. A two-tower retrieval model concatenates attribute embeddings with the learned id embedding inside each tower, so a freshly ingested item still has a usable vector from metadata alone. Hybrid designs keep the signals in separate models and switch between them.
Why it matters, and its limit
Side information is what makes cold start survivable and what keeps the long tail reachable under heavy sparsity. It is also the weaker signal: once real behaviour exists, a model leaning too hard on metadata underfits and recommends by category rather than by taste. Its quality is a practical constraint too — catalogue metadata in production is incomplete, inconsistent and frequently wrong.
Side information is any signal about users or items that does not come from the interaction matrix: item attributes, text and images, user demographics, context, or knowledge-graph relations. Because it exists even for an entity with no interactions at all, it is the standard remedy for cold start and for extreme sparsity. Models absorb it as extra feature slots in a factorization machine, or as metadata embeddings beside id embeddings in a two-tower retrieval model.
When a brand-new film lands on a streaming service, nobody has watched it yet. There is no "people who liked this also liked" to draw on. But you still know plenty about it: who directed it, what genre it is, what the poster looks like, what the blurb says.
All that extra knowledge is side information — everything the system knows that is not about people's behaviour. It is what lets a service make a sensible guess about something nobody has touched, or about a visitor who has barely browsed. Behaviour remains the stronger clue once there is any. Side information is what fills the gap until then.