Experiments on public datasets for pytorch-lifestream library

Setup and test using pipenv

# Ubuntu 18.04

sudo apt install python3.8 python3-venv
pip3 install pipenv

pipenv sync  --dev # install packages exactly as specified in Pipfile.lock
pipenv shell
pytest

# run luigi server
luigid
# check embedding validation progress at `http://localhost:8082/`

# use tensorboard for metrics exploration
tensorboard --logdir lightning_logs/ 
# check tensorboard metrics at `http://localhost:6006/`

Run scenario

We check 5 datasets as separate experiments. See README.md files in experiments folder:

• Age
• Churn
• Assess
• Retail
• Scoring
• Small demo dataset

See also additional list of experiments

Notebooks

Full scenarious are console scripts configured by hydra yaml configs. If you like jupyter notebooks you can see an example for AgePred dataset in AgePred notebooks

Results

All results are stored in */results folder.

See detailed results tables

Other experiments

• Data Fusion Contest 2024, 2-st place on the Churn Task (in Russian)
• Data Fusion Contest 2024, Ivan Alexandrov (in Russian)
• Data Fusion Contest 2022. 1-st place on the Matching Task
• Alpha BKI dataset
• American Express - Default Prediction
  • Supervised training with RNN
  • Supervised training with Transformer
  • CoLES Embedding preparation
  • CoLES Embedding usage as extra features for catboost
• Softmax loss - try CoLES with Softmax loss.
• Random features - how CoLES works with slowly changing features which helps to distinguish clients.
• Small prretrain - check the CoLES quality depends on prertain size.
• ILMC for aggregate values estimation - Imitation Learning Monte-Carlo for CLTV on Acquire Valued Shoppers Challenge dataset
• COTIC - pytorch-lifestream is used in experiment for Continuous-time convolutions model of event sequences.