Feature Combination Hybrid
Most ways of mixing recommenders run two systems and then reconcile their answers. A feature combination hybrid does something simpler, and often stronger: it keeps one model and feeds the second source in as extra input columns.
How it works
Collaborative signal is turned into features and appended to the content features a single model already consumes. In practice that means a user's interaction counts, co-occurrence statistics for an item, or a pre-trained item embedding from a factorisation model, sitting in the same feature vector as genre, brand and price. One learner — a gradient-boosted tree, an FM, a ranking network — then decides from data how much weight each source deserves, example by example. There is no blending coefficient to tune and no second model to serve.
What it buys
Folding the sources together at the input changes three things:
- A new item arrives with content features filled in and collaborative features empty, so the model degrades gracefully instead of having no opinion at all.
- The interaction between sources becomes learnable — trust co-visitation for popular items, trust the description for the long tail is a rule the model can discover rather than one you must write.
- Debugging stays in one place, because a strange ranking is explained by one model's inputs.
Why it matters
This is the shape most production ranking stages actually take: a single Wide & Deep-style network or boosted ranker eating behavioural and content features together. The cost is engineering discipline. Those collaborative features come from a feature pipeline that must compute them identically offline and at request time, or training and serving quietly disagree.
A feature combination hybrid merges recommendation sources at the input level: collaborative signals — interaction counts, co-occurrence statistics, pre-trained item embeddings — are added as features to a single model that already uses content attributes. One learner then weighs the sources from data, with no blending coefficient and no second system to serve, and a brand-new item still scores from its content features alone.
Two shop assistants could advise you. One knows what the product is; the other knows what customers like you did. You could ask both and average their answers — or you could train a single assistant who looks at both kinds of information at once.
A feature combination hybrid is the second approach. The facts about an item and the clues from other people's behaviour are handed to one system as a single list of things to consider, and it learns for itself which clue to trust in which situation. The payoff shows when a product is brand new and nobody has touched it yet: the behaviour clues are missing, the facts are still there, and the suggestion does not collapse.