dhockaday / dghl

DGHL

This repo provides an implementation of the DGHL model and produces the results for the main table presented in the paper.

Data

The datasets are not included in the repository, but are available in the following links:

NASA

SMAP and MSL datasets are available at https://s3-us-west-2.amazonaws.com/telemanom/data.zip. Labels are available at https://raw.githubusercontent.com/khundman/telemanom/master/labeled_anomalies.csv. Refer to https://github.com/khundman/telemanom for more details of these datasets.

SMD

SMD dataset is available at https://github.com/NetManAIOps/OmniAnomaly.

Environment

Run setup.sh to create a conda environment with the packages needed to run the experiments (this file does not consider different CUDA version to run in GPU. Refer to https://pytorch.org/get-started/locally/ for PyTorch installation).

Run DGHL experiment from console

To replicate the results of the paper, in particular to produce the anomaly scores for DGHL, run the `run_dghl.py' script:

python src/run_dghl.py --random_seed 1 --experiment_name 'DGHL'

The random_seed argument specifies the random seed for the model parameters initialization and Langevin Dynamics initial state. To replicate results of the paper, run with seeds 1-4. Experiment name can be used to identify different runs. This script will produce anomaly scores for the four datasets considered in the paper.

Evaluation

The evaluation functions are included in utils_evaluation.py. Use the notebook analyze_complete_run.ipynb to run the evaluation. You might need to modify the directory of the results. The notebook simple_benchmarks.ipynb runs simple benchmarks, Mean Deviation and Nearest Neighbors, on the four datasets.