Reproducing the experiments

In order to reproduce our experiments the easiest way possible, we strongly urge you to use poetry. You can install all the necessary modules via:

poetry install

Gathering the datasets

Synthetic data

Synthetic data can be generated via the following command:

poetry run python3 src/preprocessing/generate_dynamic_graphs.py --type=TYPE

TYPE can be any of the following: erdos_renyi, barabasi_albert_random, barabasi_albert_growth

Real-world data

Fetching the datasets

To store experimental data, create a new folder data as well as data/raw_datasets.

After running the experiments, your folder structure should look like this:

filtration_surfaces/
├── data/
│   ├── raw_datasets/
│   ├── preprocessed_data/
│   └── labeled_datasets/
├── src/
│   └── ...

All raw datasets shall be stored in raw_datasets. To reproduce our experiments, you can use the following command from the source directory:

wget -P data/raw_datasets https://www.chrsmrrs.com/graphkerneldatasets/infectious_ct1.zip

You can of course replace the link with a link to any other temporal dataset from TUDataset (available at https://chrsmrrs.github.io/datasets/docs/datasets/). Other temporal datasets are also supported, but we only provide a preprocessing script for the temporal datasets from TUDataset.

Preprocessing the datasets

We need to preprocess the datasets to match the dynamic graph representation that our model expects. To do so, run the following:

poetry run python3 src/preprocessing/preprocess_datasets.py --dataset=DATASET --path=PATH_TO_THE_DATASET

So, to preprocess the infectious_ct1 dataset, we would run:

poetry run python3 src/preprocessing/preprocess_datasets.py --dataset=infectious_ct1 --path=./data/raw_datasets/infectious_ct1

Training and testing our method

To train and test our method, run:

poetry run python3 src/main.py --dataset=DATASET

where DATASET is the name of the folder in data/labeled_datasets/ containing the dataset. This will construct filtration surfaces for the given dataset and store them in data/preprocessed_data/DATASET for later reuse. Afterwards, the filtration surfaces are used as an input for the random forest classifier.

If the filtration curves already exists in data/preprocessed_data/DATASET (because they were already constructed by a previous run), then the filtration curves are directly loaded.

This code adapts and extends existing code from O'Bray, Rieck and Borgwardt (2021) on filtration curves. Their code can be found at https://github.com/BorgwardtLab/filtration_curves and their corresponding paper shall be referenced as:

@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},
}