🤖 AI / ML

Collaborative Filtering

CF · Collaborative Recommendations

Definition

Collaborative filtering is a recommendation algorithm that predicts a user's preferences from the behaviour of similar users or from patterns of joint consumption between items. Unlike content-based filtering, it never looks at what a product is — only at how people interact with it.

How it works

Collaborative filtering rests on one insight: people whose past behaviour is similar tend to have
similar tastes. If users A and B bought the same ten products, then the products A bought and B did
not are a good recommendation for B.

Mathematically this is expressed through the interaction matrix:

          Item 1   Item 2   Item 3   Item 4
User A       1        1        0        1
User B       1        1        0        ?
User C       0        1        1        1

CF predicts that B will probably buy Item 4 — as A did, and B resembles A.

Two CF architectures

Memory-based CF finds similar users or items directly, through cosine similarity or Pearson
correlation. It works for small catalogues and does not scale to millions of users.

Model-based CF trains a model (matrix factorization, neural networks) that compresses the
interaction matrix into compact vectors, or latent factors. It scales to large catalogues and is
faster at inference time.

Parameter	Memory-based	Model-based
Scalability	Limited	High
Cold start	Not addressed	Partly addressed
Training speed	No training	Needs regular retraining
Explainability	High (people like you)	Low (latent factors)

Item-based CF in practice

In e-commerce, item-based CF is preferable to user-based for several reasons:

There are orders of magnitude fewer products than users, so the matrix is more manageable
Item similarity is more stable over time: coffee machine plus capsules is a durable pair
It does not require finding similar users in real time — the results are pre-computed

Tip: for a Frequently bought together block, use item-based CF over co-purchase patterns. For a
You might like block, use user-based or model-based CF across the full browsing history.

Limitations

Cold start — no history means no recommendations. Fix: content-based filtering as a fallback.
Matrix sparsity — in large catalogues most cells are empty. Fix: matrix factorization.
Popularity bias — the algorithm gravitates toward popular products and ignores the long tail.
Fix: diversification and explorative strategies.
Filter bubble — the shopper only ever sees what resembles their past. Fix: add a novelty component.

Frequently asked questions

How is user-based CF different from item-based CF?

User-based CF looks for users similar to the target user and recommends what those people liked. Item-based CF looks for products that are similar by co-consumption patterns — people who bought A also took B. In production systems item-based CF is usually preferable: it is more stable (there are fewer products than users), it scales better, and it does not suffer from drift in user interests.

What data does collaborative filtering need?

A user–item interaction matrix: views, add-to-cart events, purchases. Purchases are the strongest signal but the scarcest. Views are a weak signal but plentiful. Most systems combine them and weight the event types (purchase = 5, add to cart = 3, view = 1).

How does collaborative filtering work alongside content-based filtering?

A hybrid recommender system combines both approaches. For users with a history, CF produces precise recommendations. For new users and new products, content-based filtering handles the cold start. Most production systems are hybrids.

Gravity Field