Deep learning homework: Song hotness - Data exploration
The exercise
We decided to create a model that is able to decide whether a song will be popular or not with some accuracy. We measured the popularity of songs with YouTube view counts and we're going to try different subsets of song attributes to teach the neural network in order the get better predictions. We will also try how the prediction goes if we teach with songs of all time or songs after 2010.
Data sources
We used the Free Music Archive github repository (https://github.com/mdeff/fma) to get songs with various attributes and used YouTube to decide how popular a song is. The dataset is a dump of the Free Music Archive (FMA), an interactive library of high-quality, legal audio downloads.
All metadata and features for all tracks are distributed in fma_metadata.zip (342 MiB). The below tables can be used with pandas or any other data analysis tool. See the paper or the usage notebook for a description.
tracks.csv: per track metadata such as ID, title, artist, genres, tags and play counts, for all 106,574 tracks.genres.csv: all 163 genre IDs with their name and parent (used to infer the genre hierarchy and top-level genres).features.csv: common features extracted with librosa.echonest.csv: audio features provided by Echonest (now Spotify) for a subset of 13,129 tracks.
Because of the size of the csv-s to run the notebooks you have to download FMA_Metadata.zip from their GitHub page.
YouTube data mining
We assumed that if a song was popular at any time then it has significantly higher view count than other videos have. Our first try was to use the YouTube Data API that provides access to the properties of youtube channels, videos etc. for developers. It's basically a hidden REST API communicating with YouTube in json format. The yt_view_count.py script is out implementation using this API to search for a video and retrieve it's view count. Later turned out Google limits the usage of the API based on quota per day and we didn't have enough time for that.
We tried another approach; avoid using the YouTube API, just scrape the information from html code. The view_count_scraper.py is our implementation. The first problem was that the program was running with one thread and it was slow. It would have taken two weeks to get its job done. When we modified the code to use multithreading, YouTube detected that we are scraping their site and we got a ban. Next thing we tried was to use rotating proxy to avoid the IP ban. We successfully implemented this part, but it took so much time that it'd take months.
We returned to the first approach and successfully bypassed the quota limitation by creating new projects with new API key, or creating new google accounts. We tried to get 100.000 song's view count and ran the script in 5000 song batches not to overrun the quota. It took about one hour for each to get the view counts.
Data preparation
We used pandas python package to work with csv data and dataframes. The first problem for us was the multiindexed csv data, that was really hard to work with in pandas. We fixed the problem in excel by concatenating the multi-indexes into one single index, so a mono-indexed csv remained.
We cleaned the dataframe, deleted the unrelevant columns which had NaN or interpolated values, or it just had nothing to do with song popularity (for instance Wikipedia url link). Then we transformed a few attributes like 'genres' to categorical attributes. Kept the first 9, and named others the remaining ones. We joined the view count column retrieved by the yt_view_count script and than merge the four datas (tracks, genres, youtube_view_counts and features) to one dataframe.
Manual
First milestone:
For the homework's first milestone you need to run the data-proc/songs_data_processing.ipynb script and place the tracks3.csv, features2.csv and genres.csv to the same folder where the scripts are. You can download them from here. These csv files are nearly the same as you can download from the original dataset FMA_Metadata.zip but we have modified their indexing for avoid multi-indexing.
To run the YouTube script with YouTube Data API you have to run youtube-scripts/songs_notebook.py's yt.get_view_counts function with the appropriate parameters and modify the Google API key in youtube-scripts/yt_view_count.py.
Second milestone:
For the homework's second milestone you need to run the data-proc/songs_data_processing.ipynb script similar way to the first milestone. In our dataset we have statistics about audio tones and spectrum dispersions like chroma cens or spectral contrast. These values lie side-by-side so at this point we have not used 1D convolution because this value borders could be inside in a window. We tried some fully conncted networks, the actual can be found in the songs_data_processing.ipynb script and tried to learn the network various subsets of our training data. We have not found the best network or the best subset but the network learned most of the time.
Third milestone:
Deep Learning Documentation.pdf consist the documentation. Our final results is in the data_proc/songs_data_processing.ipynb script that you can run as described in the previous milestone. Our best model was a simple feed forward network than we used hyperparameter optimization on that. full_tracks.csv here: https://drive.google.com/open?id=1CG7zMMikkyEo9LO9Fb0JWd2OfoCeMF_Z
References
FMA: A Dataset For Music Analysis Michaël Defferrard, Kirell Benzi, Pierre Vandergheynst, Xavier Bresson