Temporal SIR-GN

This repository provides a reference implementation of Temporal SIR-GN as described in the paper:

Temporal SIR-GN: Efficient and Effective Structural Representation Learning for Temporal Graphs

Janet Layne, Justin Carpenter, Edoardo Serra, and Francesco Gullo.

The Temporal SIR-GN algorithm generates temporal structural node representations for a dynamic graph. These representations capture the evolution of each node's structure over time. Temporal SIR-GN has demonstrated superior scalability and performance on a diversity of datasets compared to existing state of the art algorithms.

Citation

If you find Temporal SIR-GN useful for your research, please cite the following paper:

@article{DBLP:journals/pvldb/LayneCSG23,
	author = {Layne, Janet, Carpenter, Justin, Serra, Edoardo, and Gullo, Francesco},
	title = {Temporal SIR-GN: Efficient and Effective Structural Representation Learning for Temporal Graphs},
	journal = {Proc. {VLDB} Endow.},
	volume = {16},
	number = {9},
	pages = {2075--2089},
	year = {2023}
}

Environment

Required packages: pandas, numpy, scikitlearn, scipy

pip install pandas numpy scipy scikit-learn

Alternatively, create an environment with these requirements:

conda create -n my-env
conda activate my-env
# If you want to install from conda-forge
conda config --env --add channels conda-forge
# The actual install command
conda install pandas numpy scipy scikit-learn

Usage

Undirected Graphs

From the command line:

python temporalSirgn.py --input --output --stop --depth --alpha --clusters  

For example:

python temporalSirgn.py --input filename --output filename --stop --depth 5 --alpha 10 --clusters 10 

Directed Graphs

From the command line:

python directed_temporalSirgn.py --input --output --stop --depth --alpha --clusters  

For example:

python directed_temporalSirgn.py --input filename --output filename --stop --depth 5 --alpha 10 --clusters 10 

Input

Temporal SIR-GN takes in a comma separated edgelist (with header) in the form of

nodeID1, nodeID2, timestamp

Output

Output will be a comma separated text file of shape n x k=(c^2+c) for an (undirected) graph with n vertices (double for directed graphs), where c is the number of clusters chosen, formatted as follows:

nodeID, dim0, dim1, dim2,...,dimk

Datasets

Undirected datasets do not have a reverse edge, however, the preprocessing from loader.py generates an adjacency list with a reverse edge. Datasets are of the form:

nodeID1, nodeID2, timestamp

with header:

src, trg, time

The table below gives the recommended hyperparameters for each dataset used in node classification tasks

Dataset	Alpha	Clusters	Depth
synth_0.0	10	10	convergence (stop = True)
synth_0.1	10	10	convergence (stop = True)
synth_0.2	10	10	convergence (stop = True)
synth_0.3	10	10	convergence (stop = True)
BrazilAir	1	10	convergence (stop = True)
EUAir	10	10	convergence (stop = True)
USAir	10	10	convergence (stop = True)
DPPIN Tarrasov	1E4	10	convergence (stop = True)
High School	1E-8	10	convergence (stop = True)
Hospital	1E5	10	convergence (stop = True)
Bitcoin OTC (directed)	1E6	11	convergence (stop = True)
GDELT	1	10	convergence (stop = True)

Note that BrazilAir, EUAir, USAir, and AS are too large for github, and can be found in the following public Google Drive:

https://drive.google.com/drive/folders/1a5uI6lIEBR3oUUU586ZQM-nJCb8pZkpg?usp=sharing

For the extremely large GDELT dataset, we refer to the AWS S3 bucket download instructions available here: https://github.com/amazon-science/tgl