mimbres/seqskip:Spotify Sequential Skip Prediction Challenge

(Since 9th Dec 2018) Paper, WSDM 2019 Cup Result

Summary

Our best submission result (aacc=0.637) with sequence learning was based on seqskip_train_seq1HL.py

• For test, seqskip_test_seq1.py
• This model consists of highway layers(or GLUs) and dilated convolution layers.
• Very similar model can be found in their encoder part of DCTTS (ICASSP 2018, H Tachibana et al.).
• Other variants of this model use various attention modules.

Other approaches (in progress):

• Multi-task learning with 1 stack generative model (aacc=0.635) that generates userlog in advance of skip-prediction is in seqskip_train_GenLog1H_Skip1eH.py
• 2 stack large generative model (not submitted) is very slow, seqskip_train_GenLog1H_Skip1eH256full.py
• Learn to compare (CVPR 2018, F Sung et al.), a relation-net based meta-leraning approach is in seqskip_train_rnbc1*.py
• Learn to compare with Some improvement is in seqskip_train_rnbc2*.py
• A naive Transformer with multi-head attention model is in seqskip_train_MH*.py
• seqskip_train_MH_seq1H_v3_dt2.py can be thought as a similar approach to SNAIL (CVPR 2018, N Mishra et al.) without their data shuffling method.
• Distillation (NIPS 2014, G Hinton et al.) approaches are in seqskip_train_T* and seqskip_train_S* for the teacher (Surprisingly, aacc>0.8 in validation, by using metadata for queries!!) and student nets, respectively. We beilieve that this approach can be an intersting issue at the workshop!
• etc.

Note that we did not use any external data nor pre-trained model.

Data split:

• 80% for training
• 20% for validation

System requirements:

• PyTorch 0.4 or 1.0
• pandas, numpy, tqdm, sklearn
• tested with Titan V or 1080ti GPU

Preprocessing:

• Please modify config_init_dataset.json to set path of original data files.
• Please run preparing_data.py(Because the data is huge, we compress it as 8-bit uint formatted np.memmap)
• Thanks to np.memmap, we can have 50Gb+ virtual memory for meta data.
• Acoustic features are loaded into physical memory (11Gb).
• spotify_data_loader.py or spotify_data_loader_v2.py is the data loader class used for training.
• Normalization:
  • Many categorical user behavior logs are decoded into one-hot vectors.
  • Number of click fwd/backwd was minmax normalized after taking logarithm.
  • We didn't make use of dates.
  • Acoustic echonest features were standardized to mean=0 and std=1.

Download pre-processed data (speed-up training with memmap)

https://storage.googleapis.com/skchang_data/seqskip/data/tr_log_memmap.dat https://storage.googleapis.com/skchang_data/seqskip/data/tr_session_split_idx.npy https://storage.googleapis.com/skchang_data/seqskip/data/track_feat.npy https://storage.googleapis.com/skchang_data/seqskip/data/ts_log_memmap.dat https://storage.googleapis.com/skchang_data/seqskip/data/ts_session_split_idx.npy (Updated, Apr 2020)

Download original dataset without login

• Training_Set_Split_Download.txt
• Sample_Submissions.tar.gz
• Training_Set.tar.gz
• Dataset Description
• Terms and Conditions
• Track_Features.tar.gz
• Test_Set.tar.gz
• Training_Set_And_Track_Features_Mini (Updated, Sep 2021)

Plots:

• plot_dataset.py can display some useful stats of Spotify dataset. You can see them in /images folder.

This repository needs clean-up!