This project provides vector embeddings for music artists. This is done by calculating a global co-occurence matrix of Spotify song recommendations to train a GloVe model.

The data for the co-occurence matrix is scraped using the official Spotify API. I have implemented two strategies for this: depth-first, where the number of recommendations for each artist is balanced, and breadth-first, where each recommended artist serves as the next search query. Which of these provides better embeddings has yet to be examined.

Data is stored in a MongoDB. Requests are optimized using multiprocessing: For every music genre a separate process is used.

To scrape your own data follow this steps:

• Create a json "credentials.json" with your Spotify API information like this:

  {
    "client_id": "...",
    "client_secret": "..."
  }
  

• Start a MongoDB instance (host "localhost", port 27017)
• Execute one of scrape-breadth-first.py or scrape-depth-first.py

An important parameter here is the number of recommendations per API request (1-100). As you can imagine, the higher the number, the worse the quality of the embeddings. However, if the number is too low, scraping takes a long time.

Once enough data is available, the co-occurence matrix can be created

python create_matrix.py

You can then train the model:

python train_glove.py

The resulting embeddings are stored as .npy file. To use them later with gensim or similar, they can be written into a text file using python make_txt.py.

I have not yet evaluated the results with any downstream task, just created some fancy visualizations. Here are some of them:

• Alternative vs. Pop Artists

• Metal vs. Classic Artists