Outlier detection under False Omission Rate control - datasets

This repository contains test datasets, their summary and visualizations used in the paper. For methods and experiments described in the paper, refer to the implementation repository.

Dataset files

Datasets used in the paper are collected in data/ directory. Each of the .csv files corresponds to a single dataset with a matching name. Each CSV file contains N columns, where first N - 2 are feature columns, and the last two are Class and BinClass columns - the former contains class assignment used in the original dataset, whereas the latter contains binarized label - 0 is treated as an outlier and 1 is treated as an inlier.

Dataset summary

All calculated dataset statistics can be found in the summary.csv file in the repository root.

Visualizations

You can find t-SNE visualization (with both original and binary classes) for each dataset in the plots/ directory, in both .png and vector .pdf formats.

Generation code

Dataset prepocessing code might be found in preprocess-all.py file. It uses datahub.io as a source for the data. the code allows for a simple adding of the new datasets or modyfing existing by defining the transformation function (preprocessing_fun) passed to the process_dataset function.