RETURNEY: Semi-supervised Prediction of the Return Time of a User

Project of "Models of Sequence Data" course, Skoltech, 2020

Problem statement

The problem is to develop a model that learns to predict users' return times from browsing histories of many other users. Some users do not come back, hence, our model needs to be able to both learn from such users' data and predict that some users will not come back during some specified prediction window.

Structure

• RMTPP - folder with our implementation of Du et al. paper
• RNNSM - folder with our implementation of Grob et al. paper
• grobformer - folder with our implementation of RNNSM model based on hawkes transformer
• data - folder with relevant datasets and helper function for their processing. dataset.py preprocessing can be reused for other datasets. Check the details in the OCON folder.
• references - folder with relevant existing implementations and their descriptions

The more detailed description of the models are available in the corresponding folders.

Usage

1. Install the requirements:

pip3 install -r requirements.txt

2. A model can be trained using train.py script:

python3 train.py 

Training parameters, including the model choice, can be changed in the file config.yaml

The key parameters are the model too train:

model = rmtpp / rnnsm / groobformer

and the metric which the best model is chosen by (if None the model is saved)

validate_by = rmse / recall / auc / none

3. The model can be tested using test.py script:

python3 test.py 

References

The project is based on the following papers:

• A Recurrent Neural Network Survival Model: Predicting Web User Return Time
  Georg L. Grob, Ângelo Cardoso, C. H. Bryan Liu, Duncan A. Little, Benjamin Paul Chamberlain, 2018

• Recurrent Marked Temporal Point Processes:Embedding Event History to Vector
  Nan Du, Hanjun Dai, R. Trivedi, U. Upadhyay, M. Gomez-Rodriguez, Le Song, 2016

• Transformer Hawkes Process
  Simiao Zuo, Haoming Jiang, Zichong Li, Tuo Zhao, Hongyuan Zha