MARD contains texts and accompanying metadata originally obtained from a much larger dataset of Amazon customer reviews, which have been enriched with music metadata from MusicBrainz, and audio descriptors from AcousticBrainz. MARD amounts to a total of 65,566 albums and 263,525 customer reviews. A breakdown of the number of albums per genre is provided here: 

Genre Amazon MusicBrainz AcousticBrainz
Alternative Rock 2,674 1,696 564
Reggae 509 260 79
Classical 10,000 2,197 587
R\&B 2,114 2,950 982
Country 2,771 1,032 424
Jazz 6,890 2,990 863
Metal 1,785 1,294 500
Pop 10,000 4,422 1701
New Age 2,656 638 155
Dance & Electronic 5,106 899 367
Rap & Hip-Hop 1,679 768 207
Latin Music 7,924 3,237 425
Rock 7,315 4,100 1482
Gospel 900 274 33
Blues 1,158 448 135
Folk 2,085 848 179
Total 66,566 28,053 8,683

 

The dataset contains 2 files and 1 folder: 

mard_metadata.json: In this file there is an entry per album. Depending on the album, the following fields may be available or not. The list of possible fields per entry: 

Source: Amazon 
amazon-id: The Amazon product id. You can visualize the album page in amazon adding this id to the following url "www.amazon.com/dp/"
artist: The artist name as it appears in Amazon
title: The album title as it appears in Amazon related:
   also bought: Other products bought by people who bought this album
   buy_after_viewing: Other products bought by people after viewing this album
price: The album price
label: The record label of the album
categories: The genre categories in Amazon
sales_rank: Ranking in the Amazon music sales rank
imUrl: Url of the album cover
artist_url: The url of the artist page in amazon. You must add "www.amazon.com" at the beginning to access this page
root-genre: The root genre category of the album, extracted from the categories field. 

Source: MusicBrainz 
artist-mbid: The MusicBrainz ID of the artist 
first-release-year: The year of first publication of the album 
release-group-mbid: The MusicBrainz ID of the release group mapped to this album title 
release-group: The MusicBrainz ID of the first release in the release-group of this album, used to extract the tracks info 
songs: List of tracks in the album 
    ftitle: Title of the track 
    fmbid: MusicBrainz recording ID of the track. Used to map with AcousticBrainz. 

acousticbrainz_descriptors/ This folder contains one file per album with the audio descriptors of its first track. In the folder, each filename has the format amazonid_mbid.json. Using the song mbid, it is possible to get all other album songs related acoustic descriptors from the AcousticBrainz web service. The number of songs with information in AcousticBrainz is constantly growing, so the numbers indicated in the Table represents the availability in the moment of the dataset creation. 

mard_reviews.json The file containing the reviews text with some metadata associated to each review. Each entry has the following fields: reviewerID, amazon-id, reviewerName, helpful, reviewText, overall, summary, unixReviewTime, reviewTime. 

Each line in these 2 files is a json dictionary. To load the file you must read a line and then eval the json expression. Python code example: 

fname = "mard_metadata.json" 
with open(fname, 'r') as f: 
    for line in f: 
        data = eval(line) 

A subset of the dataset was created for genre classification experiments. It contains 100 albums by genre from different artists, from 13 different genres. All the albums have been mapped to MusicBrainz and AcousticBrainz. It contains semantic, acoustic and sentiment features. It is stored as a dictionary, where the keys are the amazon-ids. The file is called classification_dataset.json 

We also provide all the necessary files to reproduce the experiments on genre classification in the paper referenced below. entity_features_dataset.json contains the entities and categories identified in the reviews for every album, entity_features_dataset_broader.json contains also the broader Wikipedia categories, genre_classification.py is the Python script used for the experiment. Finally, train_x.csv and test_x.csv contains the 5 different splits in the dataset used for cross validation. The genre classification dataset and evaluation scripts are available in the following GitHub repository https://github.com/sergiooramas/music-genre-classification. 

For more details on how these files were generated, we refer to the following scientific publication. We would highly appreciate if scientific publications of works partly based on the MARD dataset quote the following publication:

Oramas, S., Espinosa-Anke L., Lawlor A., Serra X., & Saggion H. (2016).  Exploring Customer Reviews for Music Genre Classification and Evolutionary Studies. 17th International Society for Music Information Retrieval Conference (ISMIR'16).

Please cite our paper in Academic Research

We would highly appreciate if scientific publications of works partly based on the MARD dataset quote the following

Oramas, S., Espinosa-Anke L., Lawlor A., Serra X., & Saggion H. (2016).  Exploring Customer Reviews for Music Genre Classification and Evolutionary Studies. 17th International Society for Music Information Retrieval Conference (ISMIR'16).

MARD dataset is available in JSON format. 

Download here the MARD dataset 

Dataset compiled by Sergio Oramas based on a previous dataset by Julien McAuley. Copyright ©2016 Music Technology Group, Universitat Pompeu Fabra. All Rights Reserved.

The MARD dataset is offered under MIT license.

Problems, positive feedback, negative feedback... it is all welcome! Please help me improve MARD by sending your feedback to: [email protected] AND [email protected]

In case of a problem report please include as many details as possible.