Filtration Curves for Graph Representation

This repository provides the code from the KDD'21 paper Filtration Curves for Graph Representation.

Dependencies

We used poetry to manage our dependencies. Once poetry is installed on your computer, navigate to the directory containing this code and type poetry install which will install all of the necessary dependencies (provided in the pyproject.toml file.

Data

We've provided sample data to work with to show how the method works out of the box, provided in the data folder. Our method works with graphs using igraph, and requires that the graphs have an edge weight (e.g., all weights in an igraph graph would be listed using the command graph.es['weight']. The BZR_MD dataset had edge weights already, and therefore we provided the original dataset; the MUTAG dataset did not have edge weights, so the data provided has edge weights added (using the Ricci curvature).

If your graphs do not have an edge weight, there are numerous ways to calculate them, which we detail in the paper. An example of how we added edge weights can be found in the preprocessing/label_edges.py file.

How to run this on your own dataset

To test out our method on your own dataset, create a directory in the data folder with your dataset name, and store each individual graph as an igraph graph (with edge weights) as its own pickle file. Then you can run the commands in the section below, replacing the name of the dataset with the name of the directory you created in the data folder.

Method and Expected Output

In our work, we used two main graph descriptor functions: one using the node label histogram and one tracking the amount of connected components. There is a file for each; but please note that the node label histogram requires that the graph has node labels.

To run the node label histogram filtration curve, navigate to the src folder and type the following command into the terminal:

$ poetry run python node_label_histogram_filtration_curve.py --dataset BZR_MD

This should return the following result in the command line: accuracy: 75.61 +- 1.13.

To run the connected components filtration curve (using the Ricci curvature), navigate to the src folder and type the following command into the terminal:

$ poetry run python connected_components_filtration_curve.py --dataset MUTAG

This should return the following result in the command line: accuracy: 87.31 +- 0.66.

Citing our work

Please use the following BibTeX citation when referencing our work:

@inproceedings{OBray21a,
    title        = {Filtration Curves for Graph Representation},
    author       = {O'Bray, Leslie and Rieck, Bastian and Borgwardt, Karsten},
    doi          = {10.1145/3447548.3467442},
    year         = 2021,
    booktitle    = {Proceedings of the 27th ACM SIGKDD International
                 Conference on Knowledge Discovery \& Data Mining~(KDD)},
    publisher    = {Association for Computing Machinery},
    address      = {New York, NY, USA},
    pubstate     = {inpress},
}