Mood-based Music Recommendations using crowdsourcing and ranking

REMARKS and TODOS (read this first!)

I have noticed that the algorithm is ineffective for a few apparent but hard to solve reasons. First, let me express what I think is effective. I think how I query Spotify to generate song lists and the stochastic nature of how I implement crowdsourcing allows for a baseline for a list of "good" songs for each specific query. However, what I do afterwords (i.e. ranking) I feel is ineffective. I utilize Spotify's given metrics to determine the rankings of each song, which is more ineffective than it is effective as it is the problem I am trying to solve. When I think of viable solutions, they seem to oppose the Problem Statement below - in other words, I think that the best solution needs some sort of personalization but not to the extent of removing the algorithm's ability to recommend a wider breadth of music. I am curious to see that applying traditional personalized content methods to crowdsourced song lists could be effective.

TODOS (maybe)

• Create a way for users to like/dislike tracks
• Use these insights to train custom model (SVM, ensemble trees, neural nets? - talking to content recommendaiton/feed ranking employees, companies are moving toward or already using deep learning methods) - based on user preferences
• Rank songs based off each user's personal model

Final remarks: I've noticed that song data is highly unavailable and pricey, so it would be hard to train models on songs data; however, I think this method may work very well for more accessible entertainment media like memes or poems.

Back to the regularly scheduled app

Table of Contents

1. App
2. Usage

App

Problem statement: it is very difficult to get music recommendations as classic algorithms (i.e. YouTube recs/Spotify 'for you') are biased toward your previous listening history and don't provide a way to appreciate a wider breadth of music.

Home Page Results for 'agitated' Example of scoring the song 'Electricity' Page showing a basic overview of how recommendations work

Usage

Clone the repository

$git clone https://github.com/neel-one/MusicMood.git

Install the necessary packages

$pip install spotipy
$pip install flask

Create three subdirectories:

artists/

moods/

songs/

(JSON files will be stored there)

(Alternatively, using a JSON object based database such as MongoDB or Firebase would be better)

Register an API key for Spotify and create 'auth.json'

{
  'username' :   placeholder,
  'id' :  placeholder,
  'secret' :   placeholder,
  'redirect' :   placeholder,
}

Optional: Pre-compute song lists for common moods

$python driver.py
$word_list.txt

Set up flask

$export FLASK_ENV=app.py

Run the program!

$flask run

Go to your favorite port to see the app (don't use this for production!)