Detecting Anomalous Event Sequences with Temporal Point Processes

Pytorch implementation of the paper "Detecting Anomalous Event Sequences with Temporal Point Processes", by Oleksandr Shchur, Ali Caner Turkmen, Tim Januschowski, Jan Gasthaus, and Stephan Günnemann, NeurIPS 2021.

Installation

1. Install the dependencies

   conda env create -f environment.yml
   

2. Activate the conda environment

   conda activate anomaly_tpp
   

3. Fix a bug in tick: There is a bug in the version of tick distributed via PyPI that makes it incompatible with scikit-learn.

   • Open the file with the bug

     nano $(dirname $(which python))/../lib/python3.9/site-packages/tick/preprocessing/longitudinal_features_product.py

   • Modify line 8 from

     from sklearn.externals.joblib import Parallel, delayed

     to

     from joblib import Parallel, delayed

   • Save and exit the file: Ctrl + X, Y, Return.

4. Install the package (this command must be run in the tpp-anomaly-detection folder)

   pip install -e .
   

5. Unzip the data

   unzip data.zip
   

Reproducing the results from the paper

• notebooks/spp_experiment.ipynb: Standard Poisson process vs. other toy TPPs (Section 6.1 in the paper).
• notebooks/multivariate_experiment.ipynb: Multivariate TPPs inspired by real-world scenarios (Section 6.2).
• notebooks/real_world_experiment.ipynb: Real-world datasets (Section 6.3).

Citation

Please cite our paper if you use the code or the datasets in your own work

@article{
    shchur2021detecting,
    title={Detecting Anomalous Event Sequences with Temporal Point Processes},
    author={Oleksandr Shchur and Ali Caner Turkmen and Tim Januschowski and Jan Gasthaus and and Stephan G\"{u}nemann},
    journal={Advances in Neural Information Processing Systems (NeurIPS)},
    year={2021},
}