Recommender Systems: Collaborative Filtering, Content-Based Methods and Cold Start
Recommenders rest on two ideas: similar users and similar items. Collaborative filtering, content-based methods, hybrids, the cold-start problem and the diversity trade-off.
A recommender answers one simple question at scale: what will this person be interested in next? The answer has two classical sources — other users who behave similarly, and other items that share features. Modern systems combine both.
Collaborative filtering
The intuition: things liked by people who liked what you liked, and which you have not seen yet, probably suit you too. To do this you build a user-item interaction matrix and factorise that sparse matrix into two low-dimensional ones. Every user and every item becomes a vector; the prediction is their dot product.
Its strength is working without knowing anything about the content. Its weakness is having nothing to say about new users and new items.
Content-based methods
Here the items’ own attributes are used: text, category, tags, imagery. You build a feature profile from what the user engaged with and recommend items with similar profiles. Embedding models are today’s standard way to measure that similarity.
This works well for new items but tends to trap users inside their own history — the filter bubble originates precisely here.
Hybrids and the two-stage architecture
| Approach | New user | New item | Serendipity |
|---|---|---|---|
| Collaborative | Weak | Weak | High |
| Content-based | Weak | Strong | Low |
| Hybrid | Good | Good | Balanced |
Cold start
What should the system recommend to a user it knows nothing about? The practical answers are simple and effective: start with popular content, ask for a few preferences at sign-up, use contextual signals such as device and locale, and weight the first session towards exploration so the system learns quickly.
Accuracy is not enough: diversity
A system that maximises click probability alone looks good briefly and then narrows: it shows the same kind of content repeatedly, the user tires, and discovery stops. Production systems therefore track three further criteria alongside accuracy — diversity, novelty and coverage.
There is also a feedback loop risk: whatever the system shows generates the data, and that data pushes the system further in the same direction. Breaking the loop requires deliberate exploration.
Frequently asked questions
Are offline metrics enough?
No. Offline ranking metrics give direction, but the real decision comes from an A/B test. Plenty of changes that look good offline lose in production.
Implicit feedback or explicit ratings?
Implicit signals such as clicks and dwell time are far more plentiful and more useful in practice. But "clicked" is not "liked", and the model should reflect that.
What are the privacy considerations?
Behavioural data is personal. Purpose limitation, retention limits and user visibility are essential; explainable recommendations build trust here.
A good recommender is not the one with the highest accuracy but the one that broadens the user over time. For more, see reranking and hybrid search and embedding models.
