Lifelong Learning of Graph Neural Networks for Open-World Node Classification

Three temporal graph datasets for node classification under distribution shift.

DBLP-Easy and DBLP-Hard are citation graph datasets. PharmaBio is a collaboration graph dataset.

Vertices are scientific publications, edges are either citations (DBLP) or at-least-one-common-author relationships (PharmaBio).

The task is to classify the vertices of the graph into the respective conference/journal venues (DBLP) or journal categories (PharmaBio). In the DBLP datasets, new classes may appear over time.

Each dataset follows the structure:

- adjlist.txt -- the graph structure encoded as adjacency lists: in each row, the first entry is the source vertex, the remaining entries are adjacent vertices

- X.npy -- numpy serialized format for node features indexed by node id corresponding to adjlist.txt

- y.npy -- numpy serialized format for node labels indexed by node id corresponding to adjlist.txt

- t.npy -- numpy serialized format for time steps indexed by node id corresponding to adjlist.txt

A paper describing our incremental training and evaluation framework is published in IJCNN 2021 (Pre-print on arXiv: https://arxiv.org/abs/2006.14422).

If you use these datasets in your research, please cite the corresponding paper:

@inproceedings{galke2021lifelong,
  author={Galke, Lukas and Franke, Benedikt and Zielke, Tobias and Scherp, Ansgar},
  booktitle={2021 International Joint Conference on Neural Networks (IJCNN)},
  title={Lifelong Learning of Graph Neural Networks for Open-World Node Classification},
  year={2021},
  volume={},
  number={},
  pages={1-8},
  doi={10.1109/IJCNN52387.2021.9533412}
}