Comment by Tallain

It's published but I would doxx myself by posting the link here -- I can email you if you'd like.

Also, full disclosure, because this was written entirely for myself, the code behind it is 100% LLM generated. I set the specs and direction but didn't touch code at all, just let agents churn while I did my day job and tested the results later.

For the recommendations:

- Seed selection is performed by building a mix of seed artists from recent listening, lifetime favorites, old obsessions, and underexplored artists.

- Seed weight combines `log(1 + play count)`, track diversity, number of listening months, and a recency decay

- Candidates are generated at the union of artists reached by "similar artist" edges through Last.fm and MusicBrainz-enriched local data (APIs are used to enrich DB initially, then after that it's all offline)

- Discovery excludes known artists, hidden and recently-reviewed artists

- Discovery requires edges with at least two distinct seeds, or one strong seed edge

- Discovery is scored by similarity (Last.fm + tag), feedback, and rediscovery thresholds. Then score is subtracted for excessive knownness within a genre, obviousness, recommendation fatigue, and shallow prior listening. The thresholds are by default but configurable:

    ```
    0.35 × calibrated Last.fm similarity
    + 0.25 × calibrated tag similarity
    + 0.10 × rediscovery
    + 0.10 × feedback
    ```

- Serendipity is additional score provided when a candidate connects several seeds that are themselves weakly connected, or are in separate connected components. This is a weak implementation of the Auralist serendipity algorithm, and will be improved on.

- Finally, diversity reranking; a greey MMR-style reranker.

    ```
    rerank score = λ × candidate score
                 − (1 − λ) × similarity to already selected results
    ```