Layer 2: personalize by re-ranking the graph per user
Two users looking at the same show saw the identical list. Layer 2 keeps a signed weight per genre and tag from a user’s own ratings and re-orders Layer 1’s existing list with it – never recomputing the graph, never scoring people twice, and never smuggling popularity in through the cold start.
Context
Layer 1 is done. The shared-people graph is global, materialized (ADR-07), and identical for everyone. It answers “shows like this” from the data and deliberately ignores popularity (ADR-05).
What it cannot do is tell two people apart. Open the same show as anyone else and the list is the same list.
The next arc is personalization (#6 for learned weights, #7 for personalized context), and it has to arrive without costing the two things that make this recommender what it is: the anti-popularity stance, and the fact that every recommendation can say why. There is a hard constraint under it too. A brand-new user has rated nothing, so whatever this is has to work from zero and get better as ratings arrive.
Decision
Layer 2 is a per-user preference profile that re-ranks Layer 1’s output.
The signal is explicit ratings plus watched, and a rating implies watched. A user rates a show 0.5 to 5.0. Submitting a rating also marks it watched, because nobody rates what they have not seen, and that holds at any score – a 1-star rating is still a watched show. Watched on its own, with no rating, is a weaker unsigned positive. These are the only inputs. Layer 2 never infers taste from something a user did not deliberately do.
It re-ranks the existing list rather than running a second engine. Layer 1 stays exactly as it is. Layer 2 takes its list and re-orders it, instead of sending the request through a separate engine that scores every show from scratch. It nudges the order of something already computed and stored, and never recomputes the graph per user. That mirrors ADR-07’s split: the store holds what is the same for everyone, the request computes what depends on the viewer.
The profile is signed weights over genres and tags. For each user the profile keeps an affinity number per genre and per tag. Rating a show moves that show’s genres and tags by a signed amount: a high rating pushes them up, a low rating pushes them down. A 1-star rating is information rather than noise – it says “less of this” – so it lowers those weights rather than being discarded. A candidate’s Layer-2 score is how well its genres and tags line up with the user’s weights.
Layer 1 owns people; Layer 2 owns genres and tags. Cast and crew already drive Layer 1’s similarity. If Layer 2 also weighted individual people, the same signal would count twice and the two layers would fight. So the division is clean. A deliberate “favors this creator” boost may arrive later, but as an explicit decision, never as an accidental second count of the people Layer 1 already uses.
Cold start encodes quality, never popularity. A new user has no ratings, so their profile starts from what correlates with high TMDb quality across the catalog: a typical-viewer prior, not a chart. This is where popularity would creep back in if it were going to, and the whole thesis of the recommender is that it does not. Until a user rates anything they mostly see Layer 1’s honest graph under a light neutral prior, and the profile personalizes away from that base as ratings arrive.
Interpretable weights, not embeddings, on purpose. Because Layer 2 weights real, nameable features, it can always explain itself: “ranked up because you rate shows with this creator or genre highly.” That explanation is the product, the context box in issue #7. A learned embedding – a black-box numeric fingerprint of a show – might rank marginally better and could not say why, which would kill the one thing that makes TVLens TVLens. Interpretability is a requirement here, not a nicety.
After Action Review
The hard part of a recommender is the design, not the code, and it is worth settling before writing anything.
Two things stuck. Keep the layers honest, so Layer 1 stays global and Layer 2 only re-ranks and nothing already working gets rebuilt. And keep it explainable, because the reason a show is recommended is the product itself – which is why nameable weights beat a smarter black box that could not account for itself.