How recommendations work

Finds users across the database whose ratings overlap with yours, measures how similar your tastes are (cosine similarity on a sparse ratings matrix), then computes a weighted average of what those users rated — giving more weight to people whose overall taste aligns more closely with yours. The result is a predicted star rating on the same 0–5 scale as Letterboxd.

Requires at least 20 rated films per person and meaningful overlap with other users' histories. When those conditions aren't met, the app moves to tier 2.

Two signals blended together:

• Affinity scoring — four signals combined into one compatibility score:
  • Genre (30%) — your average rating per TMDB genre
  • Keyword (45%) — your average rating per thematic keyword ("anti-war", "heist", "moral ambiguity") — weighted higher because keywords are more specific than genres
  • Director (15%) — your average rating across a director's films
  • Cast (10%) — your average rating across films sharing top cast members; directors carry 2× the weight of cast within this signal

  All four signals use temporal decay — ratings from 18 months ago carry half the weight of a rating from today, so the scores reflect your current taste rather than a flat average of your entire history.

• Semantic embedding similarity — when enabled, each film's plot is embedded as a vector using an AI language model. Your taste vector is the temporally-weighted average of embeddings of films you've rated above your personal mean. Candidate films are ranked by cosine similarity to that vector. Finds throughlines that genre labels miss entirely.

When both are available: final = 0.45 × affinity + 0.55 × semantic. When only affinity is available, it's used alone.

Only needs your own history — no other users required. Enable semantic matching on the Setup page.

When neither collaborative filtering nor affinity scoring can produce a result (e.g. the film has no genre data, or the user's history is brand new), the app falls back to TMDB's average audience rating. This is shown as TMDB: X.X rather than Match to be transparent that it's not personalised.