Spotify Radio

Multiple personalized radio stations on Spotify

Idea

Music streaming services provide personalized radio stations based on user preferences and music genre/style/mood/epoch. However, there is no way to define desired style/mood by yourself. Here I implemented backend (and redimentary client for a demo purposes) for alternative Spotify client with multiple personalized radio stations defined and managed by user.

Interactive docs

Implementation

The main parts of the service are:

FastAPI based webserver
radio station manager, which handles all business logic (user feedback processing, recommendations postprocessing, track to artist conversion, etc.)
vectorizer implements all logic with item vectorization, searching for items according to listening history and station style
Spotify client is responsible for (obviously) interfacing with Spotify API and could be replaced to provide the same functionality for different music streaming services

Operation details

service is completely stateless, despite cache for requests to Spotify API (so some warm-up is necessary for new replicas)
recommender is realtime, which means user feedback (listening history in this case) affects recommendations immediately
service is model agnostic, ANN index implements abstraction above algorithms (e.g. ALS, BPR, word2vec, lightFM and content-based) which could be swapped with no effort (if similar items have similar vectors and vector arithmetics works)
radio could be shared, so you can curate radio station for your followers or make station for particular event
importance of station style, positive and negative feedback could be adjusted easily and independent of each other

Model details

model recommends artists instead of tracks and tracks are sampled by heuristic due to dataset sparsity and a long tail problem distinctive to the music industry
model trained on the dataset with a user made playlists scrapped from social network
dataset was cleaned up massively because artist names are entered by users
you can find ANNOY index with item vectors and pickle with ID to name mappings in ./data

How to run it

open ./notebooks/demo.ipynb in jupyter, here you will find a demo based on rudimentary client
to make new radio station, create playlist prefixed with 'radio: ' on Spotify; playlist artists will define radio station style

How to modify it

create Spotify app account as described here
build a docker image with your Spotify settings from the root of this repo

  docker build
  --build-arg spotify_client_id=<SPOTIFY_CLIENT_ID>
  --build-arg spotify_client_secret=<SPOTIFY_CLIENT_SECRET> 
  -t spotify_radio .

run service docker run -p 8080:8080 spotify_radio localy
instantiate RadioClient without radio_host parameter