Anomaly Detection using tree-based algorithms. This repository contains code to reproduce results of figures 7 and 8 from the paper "Back to the Roots: Tree-Based Algorithms for Weakly Supervised Anomaly Detection" (arXiv link soon). Additionally, this code can be used to run other experiments by adjusting the respective parameters for the training and evaluation of several tree-based algorithms or DNNs.
IMPORTANT NOTE: If you git clone this repository, make sure to use the --recursive option in order to also correctly clone the necessary submodule!
In order to run this code, git-lfs is required. Install git-lfs with your favourite package manager, or follow the instructions e.g. on GitHub
For other recommended software packages, it is highly recommended to use a virtual environment containing the exact same packages used in our studies for optimal reproducibility. We included an environment.yml file which can be used with anaconda to install all required packages. For instructions how to install anaconda, refer to this link. Once anaconda is installed, create a new environment from the environment.yml file running the following command:
conda env create -f environment.yml --name <env_name>where <env_name> is the chosen name of the new environment.
The code can be easily run from a jupyter notebook. The notebook for reproducing results of figures 7 and 8 is called compare_models.ipynb and contains step-by-step instructions to either train all the models from scratch or load pre-trained models and evaluate them.
There is also a notebook called simple_run.ipynb which can be used to get an idea of the main functions this code provides. It contains a simple example of how to train a model and evaluate it on the test set, as well as plot a SIC curve.
The notebook tmva_comparison_study.ipynb contains code for specifically running studies using the ROOT TMVA BDT implementation.
For concrete documentation of the functions, refer to the docstrings in the code. The main code is contained in utils.py and the main plotting functions are in plot_utils.py.