playlist.append()

Spotify playlist builder via collaborative filtering recommendation system

Goal

Allow for users to create unique Spotify playlists of songs similiar to the ones they already love, simply by entering the name of an artist(s) — discovering new music in the process.

Premise

1. User inputs artist(s)
2. Recommendation system suggests playlists based on input
3. Takes random songs from a random selection of those recommended playlists
4. Formulates new playlist from those songs

File Directory

• playlist_builder.py - full code for playlist builder / recommendation system
• functions.py - just the essential functions for running the program
• main.py - main function to run the program in the terminal
• frontend.py - code for Streamlit frontend site

Prerequisites

Spotipy — Python library for the Spotify Web API

• Follow the instructions to install and get authorized before use
• Create and register your app on Spotify for Developers to obtain your Client ID and Client Secret
• We'll be using both the Client Credentials Flow and Authorization Code Flow

Data Preprocessing / Feature Engineering

• Started by scraping all playlists resulting from the queries of 126 genres from sp.recommendation_genre_seeds()
  • Can capture at most 5,000 results from each query*
• Cut that number down to 14 of what I deemed as the most popular genres after some EDA, Google searches, and my own intuition in efforts to lower the amount of data due to time constraints
• Gathered the artists that appeared in each of these playlists and took note of how many songs in the playlist were by/with that artist
• Continued to cut down amount of data to improve recommendations:
  • Eliminated artists that were in fewer than 12 playlists (mean #), and removed the playlists that those artists appeared in
  • Set maximum percentage of songs by/with an artist in each playlist at 20%, meaning each playlist had to have at least 5 different artists (got rid of playlists where greater percentages existed)
  • Removed all playlists that had less than 20 songs or greater than 500 songs
  • Ignored cover and piano versions of songs
• Ended up with 6,706 unique playlists and 17,421 unique artists

More Feature Engineering

• Created dictionary, where the 14 genres were the keys and the values were a list of playlists in that genre to use for the predictions function
• Marked each artist with a list of genres of the playlists he or she is in, ordered that list by occurences of each genre, and kept the top occuring genre
• Created second dictionary with the artist's name as the key and a tuple of his/her artist ID and top genre as the value
  • This dictionary is used to convert user input to the artist ID and recommend playlists based on that artist's top genre

Model

Collaborative Filtering with Surprise SVD

• Recommender that deals with rating data — in this case ratings = weights (percentage of songs in each playlist by/with that artist)
• Fills zeros in sparse matrix with estimated values given surrounding values

Optimal Parameters After Multiple Grid Searches:

• n_factors = 0, n_epochs = 50, lr_all = 0.009, reg_all = 0.09
• RMSE = 0.0146 (on scale of 0-1)

How It Works

• Prediction function takes in artist ID, list of playlists in that artist's top genre, and the desired # of recommendations
• Returns sorted list of recommended playlists based on actual or estimated weights
• Separate functions are used to grab the songs from those playlists, keep a random number of songs from each of them, and put them into a newly created Spotify playlist for the user

---

Demo

Created frontend site for local usage with Streamlit — an open-source Python library used to create custom web-apps for machine learning (see Demo link).