Weighted Hybrid
When you have two recommenders that are each good at something different, the simplest useful thing you can do is let them both vote — and decide how loudly each one gets to speak.
A linear blend of scores
A weighted hybrid runs several recommenders over the same candidate set and combines their scores into one, typically as a weighted sum: 0.7 times the collaborative-filtering score plus 0.3 times the content-based score. Every component scores every candidate, which is what distinguishes this from a switching hybrid, where one recommender is chosen per request, and from a cascade, where one refines another's ordering. Scores from different models are not on a common scale, so normalisation — or rank-based fusion such as reciprocal rank fusion — comes before the weighting.
Setting the weights
Weights can be fixed by hand and tuned offline, learned by fitting a small model on held-out interactions, or made dynamic: lean on content similarity while a user has almost no history, then shift toward collaborative signal as interactions accumulate. That last pattern is the standard mitigation for new-user cold start, and it is why weighted hybrids remain common long after more elaborate architectures arrived.
What to watch
Weights tuned against an offline metric can quietly favour the component that best reproduces past exposure, inflating popularity bias. And the blend is only as cheap as its slowest component — every model runs on every candidate, inside the same latency budget.
A weighted hybrid combines several recommenders by scoring every candidate with each of them and summing the scores with fixed or learned weights. Because component scores sit on different scales, they are normalised first. The weights can shift with context — favouring content-based signal for a user with little history, collaborative signal once interactions accumulate — which makes it a standard, cheap defence against cold start while keeping one ranked list.
Picture asking two friends what to watch. One knows what films are popular with people like you; the other knows what films resemble the ones you already love. Instead of picking a favourite friend, you listen to both — and give one of them a bit more say.
A weighted hybrid recommender does exactly that arithmetic. Each method gives every option a score, and the system blends those scores, deciding how much each opinion counts. It can even change its mind: trust the taste-matching friend more once it knows you well, and lean on the similar-films friend while you are still new.