Overview

This is the implementation for the EigenPulse algorithm described in the paper: "EigenPulse: Detecting Surges in LargeStreaming Graphs with Row Augmentation".

In this paper, we reasonably model the streaming graph as a row-augmented matrix, and propose, EigenPulse, to detect surges in large streaming graphs, based on the singular spectrum of the matrix. EigenPulse has the following advantages: (1) Incremental singular value decomposition: we propose an algorithm, AugSVD, which can output the spectral values of graph nodes at each time window; (2) Robust: we theoretically analyze that the robust approximation of AugSVD to batch SVD. (3) Effectiveness: EigenPulse can detect suspicious synchronized activities accurately in real-world graphs.

Requirements

This project is written in Python 3.6 We suggest recreating the experimental environment using Anaconda through the following steps.

1、Clone the project

git clone https://github.com/shenghua-liu/EigenPulse.git

2、Install the appropriate version for Anaconda from here - https://www.anaconda.com/distribution/

3、Create a new conda environment named "eigenpulse" and install requirements. comment: <> (pip install --user --requirement requirements.txt)

conda create -n eigenpulse python=3.6
conda activate eigenpulse
conda install --force-reinstall -y --name spartan -c conda-forge --file requirements.txt

Experiments

• For demo, we use BeerAdvocate dataset

python beer.py

• Input data should be in chronological sequence, like /dataset/beer/input.tensor, a preprocessed BeerAdvocate data.
• The output dir is /output/beer/.
• See beer.py for complete parameter introduction.