Track2Seq is a Deep Long Short Term Memory network implementation which can be used to generate diverse playlist continuations by predicting one track at a time. The method leverages top-k next-item probabilities to construct a list of recommendations through a semi guided prediction process. In addition the setup shows how title information can be used for playlists when no seed tracks are available.
All dependencies are in
requirements.txt and can be installed i.e. through
pip install -f requirements.txt
The network was designed to work with the Million Playlist Dataset (MPD), official website hosted at https://recsys-challenge.spotify.com.
After setting local variables in
src/config.json pre-processing and training scripts can be executed. Make sure to download the pre-computed word2vec embeddings for step 6) by Mikolov et al. Run the scripts in following order:
This will perform the following steps:
- Generating statistics for all playlists to stratify on
- Binning and stratification of playlists
- Splitting of playlists in train, development and test sets
- Bucketing of development and test sets to match challenge data
- Turning training playlists in int-sequence and filtering less frequent songs (< 5)
- Computing seed tracks for 0-seed playlists
The whole process will take a while to compute and requires sufficient memory as well as disk space. Pre-computed files as well as weights will be made available as well.
Afterwards the model can be trained by running