Hybrid Recommender

Every recommendation signal fails in its own way. Collaborative filtering is helpless with a brand-new item, a content model keeps offering more of the same, and a popularity list is identical for everyone. A hybrid recommender is the design that stops pretending one signal is enough.

The standard ways to combine

• Weighted — scores from several recommenders are blended, with weights tuned or learned.
• Switching — a rule picks one recommender per situation: attributes while an item has no interactions, collaborative evidence once it has them.
• Mixed — lists from different sources appear side by side as separate shelves.
• Feature combination and meta-level — interaction signals enter one model as extra features, or one model's learned representation becomes another's input.
• Cascade — a cheap recommender produces a candidate set and a stronger one re-ranks it.

Why it matters
The combination is not decoration; it covers each component's blind spot. Attributes carry a new item until interactions accumulate, and collaborative evidence supplies the surprise a content model cannot. The retrieve-then-rank architecture serving most production systems is itself a cascade hybrid — several retrieval sources, one ranker deciding between them. LightFM is the compact example: an embedding is the sum of feature embeddings, so one model degrades gracefully to pure factorization or pure content depending on what data exists.

What goes wrong
A blend with untuned weights lets the noisiest source dominate, and a switching rule written once quietly stops matching the catalogue it was written for.