Your music recommender uses collaborative filtering. Newly released tracks are never recommended. Why, and what do you do?

What the interviewer is testing

Whether you recognise a structural blind spot in a technique rather than treating it as a tuning problem.

Why it happens

Collaborative filtering recommends what similar listeners played. A track nobody has played has no co-occurrence signal, so it is invisible to the model — permanently, unless something else surfaces it.

This is not a tuning issue. No amount of model improvement fixes an absent input, and teams routinely spend months optimising a model whose failure mode is missing data.

For a service whose value proposition includes discovery, it is not a corner case: the tracks users would most value discovering are by definition the ones with the least listening data.

The fix: additional signals, not a better model

Content-based features from the audio itself. A model over the audio signal produces a representation for a track with zero listening history, which addresses cold start directly. This is the piece that makes day-one recommendation possible.

Text signals. Natural language processing over reviews, articles, blogs and playlist titles gives a signal for tracks people write about even before many have streamed them, and captures cultural context that behaviour alone does not.

Metadata and artist relationships, so a new release by a known artist inherits some signal.

Deliberate exploration. Reserve a fraction of recommendations for items with little data, and treat the resulting plays as the signal that bootstraps collaborative filtering. Without this, popular items get more plays and become more popular — a feedback loop that entrenches the catalogue.

Spotify's Discover Weekly is publicly described as combining collaborative filtering, NLP over text, and raw audio analysis for exactly these reasons.

What a strong answer adds

The architectural point: the difficulty is not the models but the platform — three different pipelines producing comparable representations, refreshed on a schedule, joined per user and served. That is where most of the engineering lives.

And the generalisation: for any ranking or matching system, enumerate the cases where your primary signal is absent — new items, new users, rare categories, long-tail queries — and design a distinct path for them rather than accepting degraded output.

Common weak answers

Retraining more frequently, which does not create data that does not exist. Boosting new items arbitrarily without a quality signal.