Here are a few self-hosted approaches you can experiment with to replace Plex/Jellyfin’s built-in recommendation algorithms:

1. Apache PredictionIO
   • A full ML recommendation server you can train on your own play-history data.
   • You’ll point its data import at the Plex (via Tautulli) or Jellyfin playback APIs, pick a collaborative-filtering template, let it build a model, then query its REST endpoint for personalized suggestions.

2. LensKit or Surprise (Python)
   • Both are lightweight Python toolkits for building and evaluating recommender systems.
   • You’d extract your user–item ratings (or play counts) from Plex/Jellyfin, feed them into a LensKit/Surprise script, train a CF model, then wrap the inference call in a small Flask/Gunicorn service.
   • Plex can consume the output via a webhook or the “Watchlist” plug-in; Jellyfin has a plugin framework where you can point to your new recommendation API.

3. LibRec (Java)
   • Another mature CF library.
   • Package as a microservice, import your server’s watch logs, and expose a simple “/recommend?user=XYZ” endpoint you point your media server’s UI at.

4. Community Plugins & Scripts
   – “plex-recommendations” on GitHub (Python + Pandas) that reads your Plex history and spits out top-N lists.
   – Jellyfin’s third-party “External Recommendations” plugin (still in active development) lets you call any HTTP API and show its results right in the Jellyfin UI.

Workflows typically look like this:
• Extract watch events from Plex with Tautulli (or directly from Jellyfin’s API).
• Feed those events into your chosen CF engine and build a model.
• Host that model behind a simple REST API.
• Point your media server (via plugin or webhook) at that API to display fresh movie/TV recommendations in your library.

That stack gives you full control—you can swap algorithms, tweak parameters, and retrain whenever your tastes evolve, all behind your own firewall.
Link to Article