3. Model Families

Model choice depends heavily on what data you have. If you only observe interactions, collaborative filtering is usually the first serious approach. If you also have user and item attributes, content-based or hybrid models become more useful. If the current situation matters, such as device, country, time, or within-session behavior, then contextual models become important.

3.1 Content-based filtering

Use user/item attributes and metadata.

Item vectors from text/category/tags/embeddings
User representation from demographics and/or consumed item profiles
Similarity models (cosine, k-NN) or supervised models over features

Strength:

Better cold-start for new items (and sometimes new users)

Limitation:

Limited collaborative signal; can over-specialize
Can become overly narrow if the feature space does not capture richer or emerging interests

Google’s course also emphasizes that content-based systems are often easier to explain and easier to cold-start for new items, but they tend to be weaker at serendipity than collaborative models.

Aspect	Advantages	Disadvantages
User specificity	Does not require interaction data from other users, so it can scale cleanly across many users and preserve privacy better	Quality depends heavily on having good item features and user profiles
Discovery pattern	Can serve niche items that match a user’s known interests very well	Usually expands less well beyond existing interests, so serendipity is weaker
Modeling burden	Easier to explain because recommendations can be tied back to item attributes	Requires substantial domain knowledge and hand-engineered or high-quality learned features

Content-based filtering feature matrix illustration

Image credit: Google for Developers Recommendation Systems course, CC BY 4.0.

3.2 Collaborative filtering

Use interaction patterns across all users/items.

Neighborhood methods (user-user, item-item)
Latent-factor methods (matrix factorization)

Strength:

Often strong personalization with enough interaction history

Limitation:

Cold-start if no history exists

Aspect	Advantages	Disadvantages
Feature engineering	Embeddings are learned automatically, so little prior domain knowledge is required	Harder to incorporate side features such as demographics, metadata, or context without model extensions
Discovery pattern	Can introduce serendipity because similar users can pull an item into a user’s candidate set	Suffers from cold-start for fresh users and fresh items without interactions
Practical role	Strong starting point because a feedback matrix alone can power a usable candidate generator	Production systems usually need extra machinery such as WALS projection heuristics, side-feature augmentation, or hybrid models to fill the gaps

3.3 Contextual filtering

Contextual filtering incorporates information about the current situation into the recommendation process.

Examples of context: device, country, date, time, session state, or recent action sequence
Useful when the same user may want different items under different circumstances
Often framed as next-action or next-item prediction rather than only long-run preference estimation

Contextual recommendation diagram

3.4 Hybrid models

Combine metadata with interaction learning.

Best default choice in many production systems
Handles cold-start better than pure collaborative filtering
Usually outperforms pure content-based methods once enough interactions accumulate

Content-based, collaborative filtering, and hybrid model comparison

3.5 Embedding spaces and similarity measures for candidate generation

The Google Developers course sharpens an important operational point: candidate generation is usually a nearest-neighbor search problem in an embedding space. Given a query embedding $q$ and item embedding $x$ , the retrieval stage depends heavily on the similarity measure you choose.

Common choices are:

$s_{\cos} (q, x) = \frac{⟨ q, x ⟩}{‖ q ‖_{2} ‖ x ‖_{2}}$

$s_{dot} (q, x) = ⟨ q, x ⟩$

$d_{L 2} (q, x) = ‖ q - x ‖_{2}$

If the embeddings are normalized, cosine, dot product, and squared Euclidean distance induce closely related rankings. Without normalization, however, they behave differently:

Dot product favors larger embedding norms, which often correlates with popular or frequent items
Cosine focuses more on angular alignment, which can be better for semantic similarity
Euclidean distance emphasizes physical closeness in the embedding space

Google also suggests a useful interpolation between pure cosine and pure dot product:

$s_{α} (q, x) = ‖ q ‖_{2}^{α} ‖ x ‖_{2}^{α} \cos (q, x), α \in (0, 1)$

This lets you keep some popularity signal without letting large-norm items dominate retrieval.

Candidate retrieval in embedding space

Image credit: Google for Developers Recommendation Systems course, CC BY 4.0.

Different similarity choices induce different rankings

Image credit: Google for Developers Recommendation Systems course, CC BY 4.0.

Last updated on Tue, Mar 10, 2026