• This repository provides codes for bulding a MRPG and a distance-based outlier detection algorithm on a MRPG.
• Our algorithm supports metric space.
  • Our codes implement L2 (Euclidean distance), L1 (Manhattan distance), Jaccard distance, Edit distance, angular distance, and L4 distance by default.
  • The other distance functions are free to add.
• The details about our algorithm can be seen from our SIGMOD2021 paper, Fast and Exact Outlier Detection in Metric Spaces: A Proximity Graph-based Approach.

• Linux OS (Ubuntu).
  • The others have not been tested.
• g++ 7.4.0 (or higher version) and Openmp.

• Before running our DOD algorithm, build an MRPG.
• Parameter configuration can be done via txt files in parameter directory.
• Data files have to be at dataset directory.
  • You can implement data input in as you like manner at input_data() function in data.hpp.
  • Now dataset directory contains a dummy file only.

• Create result/graph directory.
• Compile: g++ -O3 -o mrpg.out main.cpp -std=c++11 -fopenmp.
• Run: ./mrpg.out.

• Create result directory.
• Compile: g++ -O3 -o greedy-pivot.out main.cpp -std=c++11 -fopenmp.
• Run: ./greedy-pivot.out.
• If you test a low-dimensional dataset, you may enable VP-tree based verification (by setting mode = 1 in main.cpp).
  • By default, verification is done by a linear scan.

If you use our implementation, please cite the following paper.

@inproceedings{amagata2021dod,  
    title={Fast and Exact Outlier Detection in Metric Spaces: A Proximity Graph-based Approach},  
    author={Amagata, Daichi and Onizuka, Makoto and Hara, Takahiro},  
    booktitle={SIGMOD},  
    pages={36--48},  
    year={2021}  
}

Copyright (c) 2020 Daichi Amagata
This software is released under the MIT license.