song2vec

This repository contains the code for a skip-gram model for learning distributed representations of songs based on the context in which they appear in a playlist.

Motivation

This project was based on a 2017 talk by Spotify ML on using AI to build predictive platforms. One of the topics in the talk revolved around leveraging NLP research on word embeddings (word2vec) to model track similarity. By modeling track similarity, a user can be recommended new and similar tracks based on the tracks they have listened to in the past.

Main idea: Based on the assumption that users tend to listen to similar tracks in a sequence, one could build a word2vec model that learns a vector embedding for any given track, using the tracks that appear before and after it in different playlists. Once trained, the embedding weights would be organized in such a way that similar tracks would have similar vectors.

Approach

Zenodo's Spotify Playlists Dataset contains 1.2GB of playlists along with the tracklist for each playlist. I extracted a small fraction of the dataset, then built and trained a skip-gram model with negative sampling using the tf.keras library.

Results

The model seems to have not only learned similar vectors for tracks that appear close to one another in a given playlist, but also managed to learn some interesting relationships between songs that appear nowhere near each other in the dataset.

Below is an example showing songs similar to John Mayer's cover of Tom Petty's song, "Free Fallin'".

top 10 tracks most similar to 'john mayer - free fallin' - live at the nokia theatre' (pos. 4602):
- (sim. 0.562): 'the martini's - free' (pos. 4600)
- (sim. 0.446): 'the cranberries - free to decide' (pos. 4605)
- (sim. 0.389): 'säkert! - fredrik' (pos. 4599)
- (sim. 0.388): 'rascals - freakbeat phantom' (pos. 4597)
- (sim. 0.388): 'tom petty - free fallin'' (pos. 206)                     <---- original track
- (sim. 0.377): 'cary brothers - free like you make me' (pos. 4603)
- (sim. 0.350): 'robot koch - nitesky feat. john lamonica' (pos. 6755)
- (sim. 0.331): 'håkan hellström - brännö serenad' (pos. 3480)
- (sim. 0.330): 'justice - the party' (pos. 264)
- (sim. 0.323): 'muse - resistance' (pos. 7376)

As can be seen above, the model was able to learn the similarity between John Mayer's cover and Tom Petty's original version, even though both tracks do not appear close to each other in the dataset. John Mayer's cover appears in position 4602 and Tom Petty's original appears in position 206.

Project directory structure

The project is organized in such a way that experiments can be run for different hyperparameter settings. Results and models from each experiment are stored in the experiments directory.

.
└── experiments/                # experiments (processed data + models)
    └── default_hyperparams/    # experiment results for default hyperparameters
├── notebooks/                  # jupyter notebooks for inference
├── raw_data/                   # part of the original Zenodo dataset
├── src/                        # cleaning, preprocessing, and training scripts
├── run_experiment.py           # script for running an end-to-end experiment
├── requirements.txt
└── README.md

Running the project

Environment set-up

1. Clone the project
2. Create and activate a new conda environment
3. Run pip install -r requirements.txt

Make inferences using the trained model

To visualize the similarities learned by the trained model, open notebooks/Inference.ipynb.

Run an experiment

The script run_experiment.py enables you to build and run experiments.

Below is an example of an experiment.

python run_experiment.py large_wsz --window_size 4

To view the list of parameters, run python run_experiment.py -h.

usage: run_experiment.py [-h] [--csv_path CSV_PATH] [-nr NUM_ROWS]
                         [-wsz WINDOW_SIZE] [-emb EMBEDDING_DIM] [-e EPOCHS]
                         [-lr LEARNING_RATE] [-rs RANDOM_SEED]
                         experiment_name

positional arguments:
  experiment_name       name of experiment

optional arguments:
  -h, --help            show this help message and exit
  --csv_path CSV_PATH   path to raw .csv dataset (defaults to
                        raw_data/spotify_playlists.csv)
  -nr NUM_ROWS, --num_rows NUM_ROWS
                        number of rows to extract (-1 for whole dataset)
                        (defaults to 10000)
  -wsz WINDOW_SIZE, --window_size WINDOW_SIZE
                        (half) window size for skip-gram contexts (defaults to
                        2)
  -emb EMBEDDING_DIM, --embedding_dim EMBEDDING_DIM
                        word2vec embedding dimension size (defaults to 100)
  -e EPOCHS, --epochs EPOCHS
                        number of training epochs (defaults to 20)
  -lr LEARNING_RATE, --learning_rate LEARNING_RATE
                        learning rate for Adam optimizer (defaults to 0.001)
  -rs RANDOM_SEED, --random_seed RANDOM_SEED
                        random seed for skip-gram generation and training
                        (defaults to 42)

Once run, the relevant files (including the model, embeddings, and tokenizer) will be saved in the experiments/{experiment_name} directory. To make inferences using this new experiment's results, simply open notebooks/Inference.ipynb and change the experiment_name variable in the notebook.

References

• Spotify Playlists Dataset by Zenodo
• Applying word2vec to Recommenders and Advertising by Chris McCormick
• Using Word2vec for Music Recommendations by Ramzi Karam
• Machine Learning & Big Data for Music Discovery presented by Spotify by Vidhya Murali and Ching-Wei Chen
• A hands-on intuitive approach to Deep Learning Methods for Text Data — Word2Vec, GloVe and FastText by Dipanjan Sarkar