Machine Learning approach to the Million Songs dataset
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Type Name Latest commit message Commit time
Failed to load latest commit information.
mxm_dataset_train.txt Final NNet for Clustering Dec 8, 2017 updated comments Dec 7, 2017

Million Songs Data Set Analysis!

This project depends on the Million Songs Dataset available at:

Initial experiments are done with the smaller, experimental subset provided at:

Contributers: Aumit Leon, Mariana Echeverria

Experiments and Results

To view our experiments and results, check out our wiki:

Directory Overview

Once you download the dataset, you'll notice that the file structure is as follows:


The data directory has subdirectories that act like volumes-- if you go deep enough you'll find the H5 files that correspond to each song.

Converting the data to a usable format

The data is given to us in HD5 format (

HD5 files are binary files, so they are not very useful to us as they are given. In order to extract data from the h5 files, use

The million song dataset provides python wrappers within that can be used to recursively loop through each subdirectory and h5 file to extract certain features of the data. will visit every subdirectory (starting from the path you give indir), and will create a CSV of the data extracted from each h5 file. You don't need to put this script any place special, just be sure to provide it a proper path for indir. The output.csv file will be created in the same directory as this python script, so be sure not to commit that CSV file to Git :)

As far as I can tell, each h5 file corresponds to one song. THat might not be true of every h5 file, maybe there is a way we can verify this?