Content-based filtering can score a new item’s features against a known profile before that item receives any interactions.
Summary
Can a telescope guide with no interactions outrank a popular drama using this user’s history? Use features ordered as science, drama and short. The liked vectors are and .
The mechanism
Their mean is , with squared length . For nonzero vectors, cosine is the dot product divided by both lengths. The guide scores . The drama scores . Neither candidate’s interaction count enters this runtime calculation.
More generally, use for nonnegative weights with positive total. Define feature version, metric, recency weights and eligibility exclusions separately. Zero profiles or zero candidate vectors need an explicit fallback; their cosine is undefined. Learned encoders may still have used cross-user evidence during training.
Learning objective
Shared content features let an unseen item match a user's interests
Derivation
Can swapping the item coordinates change the match score?
Change the evidence or list policy
A user has no history
New task: recommend before any liked-item profile exists.
Picture and change: an empty mean is not the profile above. Request an optional genre preference or use a separately evaluated contextual prior.
Check: explicit drama preference gives . Drama now scores one; the telescope guide scores zero. Absent that evidence, do not pretend a zero profile has meaningful cosine scores.
Use when: the missing evidence belongs to the user, not merely a new item.
Avoid a duplicate list
New task: return two items from different categories.
Picture and change: two identical guides both score ; drama scores . Unconstrained top-two returns both guides, totaling .
Check: a one-per-category constraint selects a guide and drama, totaling . The reduced similarity sum buys the stated diversity constraint, not guaranteed user benefit.
Use when: repeated near-duplicates violate the list requirement; evaluate discovery and useful outcomes separately.
Reconstruct the information path
Can a content item tower prove that the entire recommender uses no collaborative information?
Reasoned answer
No. Inspect user inputs, training objectives and encoder provenance. Two items with identical features remain indistinguishable to this score even when their quality differs. A similarity contribution is not a causal explanation of satisfaction.
Content-based vs collaborative filtering
| Mechanism | Required evidence | Test separately |
|---|---|---|
| Content score | compatible item features and user/context profile | new-item feature coverage |
| Collaborative score | shared interaction structure or learned behavioral factors | interaction sparsity and exposure |
| Hybrid | explicitly combined feature and behavioral paths | cold/warm quality and resource costs |
Strengths and weaknesses
Features enable scoring before item interactions, but missing attributes and domain shifts can defeat that benefit. Concentrated recommendations establish overspecialization, not a causal filter-bubble conclusion. Neither approach has universal accuracy, niche-user or discovery superiority.
Hybrid systems (what’s actually deployed)
Mixtures, feature-rich two-tower models and graph side-information can combine evidence. Specify fusion weights and compare baselines on cold/warm, diversity and eligibility slices. Do not assume a mandatory age-based handoff to collaborative filtering.
What an interviewer expects you to say
Trace profile construction, candidate information, score, exclusions and evidence gaps before choosing a hybrid.
Common confusions
Item cold-start is not user cold-start. A named architecture does not specify its actual information sources.
Related: Matrix factorization for recsys, Two-tower retrieval, How would you do cold-start for a new user?, Knowledge-graph embeddings, TF-IDF and BM25.