R-Linear-search

Implementation of the BISECT algorithm for Hidden Outlier generation.

Source Code (src)

Source Code can be found in the src/ directory. Directory HOGen/ contains all of the necessary code to run the Hidden Outlier generators. We implemented both BISECT and HIDDEN [1]. Directory ODM/ contains all the necessary code about the adversaries. This includes fitting, inference, and handling Directory registry_extras/ contains all the auxiliary files used for the implementation. It includes some new methods for set class objects and the dataset handler for the method.

Installation Guide

To install, it suffices to add all of the packages from the 'R' files included via Rstudio package handler. If one wants to run Python scripts (necessary to use any adversary beyond the Mahalanobis distance) one needs to install the following dependencies inside a virtual environment named hidden_out:

• Python 3.10
• NumPY 1.23
• Pandas 1.5.2
• Pillow 9.3
• PyOD 1.0.7
• TensorFLOW 2.11
• Scikit-Learn 1.2.0
• Scipy 1.9.3
• Seaborn 0.12.2
• Numba 0.56.4
• Jupyter 1.0.0

We included the file requirements.txt including all dependencies from our virtual environment. After cloning the repository, simply execute: > pip install -r requirements.txt

User Guide

An example of use to generate hidden outliers using this code is due, and will be included in the file example.R. The directory experiments/ contains all of the necesary files to reproduce the experiments from [2].

References

[1] Steinbuss, G., Böhm, K. Hiding outliers in high-dimensional data spaces. Int J Data Sci Anal 4, 173–189 (2017). [2] Anonimized, Efficient Generation of Hidden Outliers for Improved Outlier Detection.