You already know: embeddings from co-occurrence (D4), implicit feedback (D7), approximate neighbour search (D4) and, from the video feed case, that a good slate is more than the top items by score. This case adds a weekly, batch-generated product: the goal is discovery (tracks new to this listener that they will love), feedback is implicit and noisy, every new track starts with no listening history, and a playlist is judged as a whole. The design questions: what counts as discovery, how to represent tracks and tastes, how to choose and order thirty tracks, and how to run a batch that must never miss Monday.
How to read this lesson. No code, on purpose. Numbers marked assumed are assumptions you state out loud.
1. The prompt, and the questions that change the design
A mid-level answer: "collaborative filtering, take the top 30 predicted tracks". A staff candidate first asks what discovery means (new track? new artist? how unfamiliar?), how big and how new the catalogue is, what feedback signals exist and how noisy they are, and what constraints apply (licensing by country, explicit content settings, artist limits). Those answers produce a filter on what the listener already knows, audio embeddings for new tracks, saves and skips as labels with context, and hard rights filters, plus a playlist designed as a whole.
Assumptions you state: