Allensmile / playlist-to-vec

An artist recommendation engine, from feeding Spotify playlists through word2vec

artistrecs

A similar artist recommendation engine powered by Spotify playlists and word2vec.

This proof of concept was inspired by two pieces and my own longstanding belief that the transitions between songs in playlists, when given enough, are valuable insights.

Some bathroom reading:

• Playlist Harvesting, by Stephen Phillips
• Distributional Similarity Music Recommendations Versus Spotify: A Comparison Based on User Evaluation, by Nevyana Boycheva

Also, quick reminder: this is a proof of concept! It's working and pretty cool, but that doesn't mean the tools are complete, the project isn't without layout or design decision kinks, or that things won't change. In fact, plan on things changing for as long as this message is here.

---

This application consists of two major components:

• A Celery-backed extraction setup for ingesting playlists from Spotify en masse. Celery workers are responsible for importing playlists and extracting artist names.
• Helper scripts for training and querying data extracted from said playlists

NOTE - project layout will be changing shortly. See TODO for what's up.

Setup

Before getting started with this project, please ensure you have the following installed:

• a C compiler
• Redis (up and running)
• word2vec bindings

Please also have your OAuth client ID and secret ready from the Spotify application you wish to use. If you need to register a Spotify app, do so here before continuing.

Install word2vec bindings

OSX using Homebrew

$ brew install --HEAD homebrew/head-only/word2vec

Setup

Install requirements

Check out this repo into a virtualenv, and then install its Python requirements using pip:

$ pip install -r requirements.txt

Set up environment variables

This application's configuration data is taken from environment variables.

• SPOTIFY_CLIENT_ID: Spotify OAuth client id
• SPOTIFY_CLIENT_SECRET: Spotify OAuth client secret
• ARTISTRECS_BROKER_URL: Celery broker URL. Defaults to localhost:6379. Read more.
• ARTISTRECS_RESULT_BACKEND: Celery result backend URL. Defaults to localhost:6379 Read more

How you set these variables is up to you.

TIP: envdir is great for this.

Workflow and output

Extraction workflow

First, a quick overview of what the Celery workers are doing, from start to finish:

1. playlist_generator task receives a request to search Spotify's playlists for a specific term. For each playlist found, playlist_generator creates a new task: resolve_playlist.

   If recycle is enabled, playlist_generator will respawn itself to collect more information. This is controlled by the max_recycles parameter - if left undefined, playlist_generator will collect all playlists available for the given term.

2. resolve_playlist receives a username and playlist id and fetches the playlist's tracks. It then compiles into a list each artist name from each track in the playlist [1]. Once all names have been collected, resolve_playlist hands off the list to export_artist_sentence_from_playlist.

3. export_artist_sentence_from_playlist accepts a user id, playlist id, and a list of artist names. It JSON-encodes them and then writes the object to a file.

   This task is run in a separate queue to avoid overlapping file writes.

Once all jobs have been processed, you should have a text file ready to be processed by the parser.py script.

[1] If the artist name about to be collected is also the last entry in the current list, it is ignored. This is a naive way of protecting against entire albums influencing transitional frequencies.

Right now, there are many moving pieces. As this project matures past "proof of concept", noise will be reduced.

Output format

word2vec works by analyzing sentences. That's a gross generalization. Here's something from the official docs:

The word2vec tool takes a text corpus as input and produces the word vectors as output. It first constructs a vocabulary from the training text data and then learns vector representation of words. The resulting word vector file can be used as features in many natural language processing and machine learning applications.

gensim, the word2vec implementation used here, expects sentences to be delivered as lists of words. In our particular use case, we're constructing sentences from artist names. Wild, right?

At the end of a run thorugh the extraction workflow described above, you'll have a file whose every line has a sentence constructed from artists in a playlist. The format for this output is:

{
    "playlist_id": "<spotify playlist id>",
    "user_id": "<spotify user id>",
    "sentence": [
        "<artist name>",
        "<artist name>",
        ...
    ]
}

This file can be found at SENTENCE_OUTPUT_PATH.

Running

Once you've completed the above, you're ready to begin.

Start Redis

If redis-server is not already running in the background, fire it up.

$ redis-server

Start extraction queues

In one shell inside your virtualenv, fire up the extraction queue. This queue is responsible for the tasks concerned with querying Spotify and compiling artist names from tracks inside of playlists.

$ celery worker -A artistrecs -l info -Q extraction

In another shell, start the writer queue with 1 worker. This is a sloppy workaround to prevent overlapping file writes by unlocked append access to a file.

$ celery worker -A artistrecs -l info -Q writer

Insert a task

To query Spotify for playlists, insert a task using insert_task.py. This helper script accepts two parameters, expecting one of them:

$ python insert_task.py -t <term>

See insert_task.py's --help print-out for extended usage details.

Parsing output

Data extracted from playlists may be parsed using parser.py.

$ python parser.py -i <path to output file> -t <artist name>

This will emit a JSON-encoded ranked list of similar artists. Make sure that the artist name you're checking was, in fact, a member of at least a few of the playlists brought in during extraction.

See --help for extended usage details.

TODO

• Add setup.py, which should also install helper scripts as console scripts
• Parser should default to using SENTENCE_OUTPUT_PATH env var when -i is not given.

Issues

Want to contribute? File an issue or a pull request.