playlist-cleaning

A cnn-based model for playlist cleaning and reranking integrated with the structure of super-resolution neural network.

This is one of my three projects during the internship at KKBOX. Note that although there is a Python file related to seq2seq model, this project had not supported the seq2seq model after an early experiment which shows that its performance is much worse than the cnn-based model.

Dependencies

• python3
• TensorFlow >= 1.2
• numpy
• tqdm

$ pip3 install -r requirements.txt
# to install TensorFlow, you can refer to https://www.tensorflow.org/install/

Files you should prepare

data/raw/x.txt

This file contains raw playlists.

[date_of_created_playlist] [song_id1] [song_id2] ....

data/raw/y.txt

This file contains rereanked playlists corresponding to data/raw/x.txt line by line.

[date_of_created_playlist] [song_id1] [song_id2] ....

Usage

Prepare data

$ ./prepare_data.sh

Train

$ python3 main.py --nn cnn --mode train

Valid

$ python3 main.py --nn cnn --mode valid

Create default testing set

$ ./create_default_test_data.sh

Test

$ python3 main.py --nn cnn --mode test