Collaborative filtering recommends items to a user based on the behavior of other users, not on any understanding of what the items actually are. If people who liked what you liked also liked something else, the system suggests it to you. No content analysis required — just a large matrix of who interacted with what, and a method for finding patterns in it.
What changed in 2026
- Neural collaborative filtering is now the default for large catalogs, with embedding-based approaches largely replacing raw matrix factorization in new builds, though matrix factorization remains a strong, cheap baseline.
- Graph-based collaborative filtering gained traction for catalogs where relationships between items matter beyond simple co-purchase, such as media platforms with franchise or series structure.
- Hybrid pipelines that pair collaborative filtering with LLM-derived content embeddings became a common way to soften the cold-start weakness without abandoning the collaborative signal entirely.
User-based vs item-based filtering
User-based filtering finds users similar to you and recommends what they liked. It is intuitive but has a real weakness: user preferences and the set of active users both shift constantly, so similarity computed today can be stale tomorrow, and recomputing it at scale is expensive.
Item-based filtering instead finds items similar to items you already liked, based on co-interaction patterns. Item relationships are far more stable over time than user relationships — a co-purchase pattern between two products does not change much week to week — which is why item-based approaches scale better and dominate in production.
How matrix factorization works, briefly
Represent user-item interactions as a large, sparse matrix (most users have not interacted with most items). Matrix factorization decomposes that matrix into two smaller dense matrices — a user-factor matrix and an item-factor matrix — such that multiplying them approximately reconstructs the original interactions. The learned factors are effectively compressed embeddings capturing latent taste dimensions the model discovers on its own, not ones a human labeled. This is the direct conceptual ancestor of the embedding-based neural methods used today.
Collaborative filtering methods compared
| Method |
How it works |
Best fit |
Main weakness |
| User-based CF |
Similarity between users |
Small, stable user bases |
Expensive to recompute; drifts fast |
| Item-based CF |
Similarity between items |
Large catalogs, e-commerce |
Needs meaningful interaction volume per item |
| Matrix factorization |
Latent factor decomposition |
General-purpose, strong baseline |
Cold start; no side information used |
| Neural CF / two-tower |
Learned embeddings, deep interactions |
Large-scale production systems |
More infrastructure and tuning required |
When collaborative filtering is the wrong tool
If your catalog is small, changes slowly, or interaction volume per item is thin, collaborative filtering will not find reliable patterns — the matrix stays too sparse. In these cases, a content-based or rule-based approach (recency, popularity, curated similarity) typically outperforms it. Collaborative filtering earns its complexity at scale, not before.
Common pitfalls
- Treating collaborative filtering as cold-start-proof. It is not — see the cold start problem for how teams actually work around this. New users and new items have zero interaction signal by definition, and the method has nothing to work with until that changes.
- Ignoring popularity bias. Collaborative filtering naturally amplifies already-popular items because they have the most interaction data, which can crowd out relevant long-tail items unless you explicitly correct for it.
- Skipping implicit feedback. Explicit ratings are sparse and biased toward strong opinions. Implicit signals — clicks, dwell time, purchases — are noisier but far more abundant, and most production systems lean on them heavily.
FAQ
Is collaborative filtering still relevant with LLM-based recommendations?
Yes. LLM-derived signals are useful for content understanding and cold start, but collaborative filtering remains the strongest signal for capturing taste patterns from actual behavior, and most systems combine both.
What is the difference between explicit and implicit feedback in collaborative filtering?
Explicit feedback is a direct rating (a star rating, a thumbs up). Implicit feedback is inferred from behavior (a click, a purchase, time spent). Implicit feedback is noisier per data point but usually far more abundant.
Does collaborative filtering need any information about the items themselves?
No — that is its defining trait. Pure collaborative filtering uses only interaction patterns, not item metadata. That is also why it struggles with cold start.
Where to go next