Awesome

Playlist Generation Based on Graph Theory

This project is part of "A Network Tour of Data Science" course.

The objective of this project is to generate a playlist that can bring the listener from his actual mood (e.g. angry or sad) to a mood he would like to be in (e.g. relaxed or happy). Through the light of networks, it is possible to examine how music tracks relate to each other. Can we build a playlist that links two very different tracks through a smooth transition? To answer this question, we will build a similarity graph between music tracks and choose a smooth path taking into account user's preference.

Usage

To start with the project, you can make a new environment and first make sure you have the following libraries installed and ready:

• Pandas
• Scikit-Learn
• Numpy
• Scipy
• NetworkX
• Seaborn
• MatplotLib
• Librosa
• Spotipy
• tqdm
• multiprocessing
• urllib
• json

Dataset

To start, you can clone the repository, then download the following zip file and add it to the 'data' folder: Download the Prepared dataset (download link)

• The initial User’s History was collected from (Million Songs Dataset): Million Songs Dataset

Notebooks for Preparing the main dataset

• Audio Features dataset were constructed and added using Spotify API: detailed codes can be found in Collecting and Preproccessing Data
• MetaData were dataset constructed and added using Spotify API: detailed codes can be found in Collecting Meta Data
• Features dataset were computed and added using Librosa Library: detailed codes can be found in Collecting and Preproccessing Data (Declaration: the functions for calculating the features were inspired from FMA dataset created by Michaël Defferrard, Kirell Benzi, Pierre Vandergheynst, and Xavier Bresson https://github.com/mdeff/fma/blob/master/features.py

Main Parts of the Project

For detailed information about the project, you may read our report:

• Report.pdf

For the detailed codes about the main part of the project, you may refer the following notebooks:

• Creating Similarity Graph

• Generating Playlist

• Generating Files for Gephi Visualizations

Other useful information about key parameters in our algorithm:

1. Seed/end song selection

If the user wants the algorithm to define the seed(start) song or end song according to his/her mood selections, he/she should assign to the variables end_song/seed_song the value returned by the function "song_selection(Danceability,Energy,Valence)" where the parameters represent the each mood dimension and should be between 0 and 1. If the user has a particular seed/end song that he wants to use, he/she can assign to the variables seed_song/end_song the corresponding node number of the desired songs (index of the adjacency matrix).

2. Cutoff (playlist length)

If the user wants the playlist generated to have a specific lenght, he/she should adjust the variable "cutoff" in the network X method all_simple_paths. The playlist generated will have a maximum lenght of cutoff+1. Be aware that for comptational time reasons, cutoff should not be too much bigger (+3 max) than the shortest path lenght (otherwise to many paths are generated).

3. Lambda coefficient

This factor can be adjusted depending on how much the user wants the songs in the playlist to take into account his taste.