Music4All is a dataset for music information retrieval and music recommendation research

music4all_contrib

Written by Keunwoo Choi

This is NOT official music4all dataset repository.

Music4All is a dataset for music information retrieval and music recommendation research. It comes with a user-based split for recommendation research. However, there's no official track-level split so I decided to share my own one.

There are several types of attributes and labels in the dataset. I am mainly interested in labels -- genres, tags, and languages.

• Language: 46-class single-label classification
• Tag: multi-label, 19,541 classes
• Genre: multi-label, 853 classes

I used scikit-multilearn for multi-label stratification and scikit-learn for single-label stratification.

Training-Testing Split

I made two splits -- split-all_labels and split-language_only. Each has only one primary 80:20 split.

split-all_labels

Split is here (text files of the track indices)

This split is a stratification based on all the labels (language, tag, genre). I merged all the labels (19541 + 853 + 46 = 20440) and performed stratification using scikit-multilearn. The result is as below (up to top 2000 labels only). See the distributions of training set (top) and testing set (bottom).

Out of 22400 labels, I recommend to consider top-500 labels (13 language, 321 tag, 166 genre labels) due to the data scarcity. (More details at in the notebook)

split-language_only

Split is here (text files of the track indices)

This split is a stratification based on language labels only. It is single-label so I used sklearn.model_selection.StratifiedKFold. See the distributions of training set (top) and testing set (bottom).

Note that in the testing set, there is no item with {'hi', 'lv', 'bn', 'bg', 'sk', 'pa', 'he'}.

Few more plots

Occurrence counts of tag labels

Occurrence counts of language labels

Same, but with log y-axis:

Occurrence counts of genre labels

Notes

• Seems like some of the tags are just identical to genre labels (e.g., the occurrence counts of the top 2 genres, rock and pop, are equal to those of the top 2 tags, rock and pop.)
• The stratification process and other analysis is in notebook.
  • You'll need to download the dataset and put csv files to id_csvs/ to run the notebook.