Visual Graph Arena Dataset

Dataset Details

Visual Graph Arena dataset contains 3 main tasks each with two sub-tasks (more details in Tasks).

The directory structure of the Visual Graph Arena (VGA) dataset can be accessed here. Also, an interactive version of the directory structure is available below.

Graph Arena(3)
- isomorphism(2)
  - easy(2)
    - inputs(2)
      - train (140,000 samples)
      - test (15,718 samples)
    - labels(2)
      - train_labels.txt
      - test_labels.txt
  - hard(2)
    - inputs(2)
      - train (127,374 samples)
      - test (14,298 samples)
    - labels(2)
      - train_labels.txt
      - test_labels.txt
- path(2)
  - hamiltonian(2)
    - inputs(2)
      - train (25,000 samples)
      - test(2)
        kawai (2,480 samples)
        planar (2,480 samples)
    - labels(2)
      - train_labels.txt
      - test_labels.txt
  - shortest(2)
    - inputs(2)
      - train (80,000 samples)
      - test(3)
        random (8,672 samples)
        kawai (8,672 samples)
        planar (8,672 samples)
    - labels(2)
      - train_labels.txt
      - test_labels.txt
- cycle(2)
  - hamiltonian(2)
    - inputs(2)
      - train (69,935 samples)
      - test(2)
        kawai (7,740 samples)
        planar (7,740 samples)
    - labels(2)
      - train_labels.txt
      - test_labels.txt
  - cordless(2)
    - inputs(2)
      - train (80,000 samples)
      - test(2)
        kawai (6,484 samples)
        planar (6,484 samples)
    - labels(2)
      - train_labels.txt
      - test_labels.txt

The graphs in the datasets consist of 8-9 nodes. The training and test sets for each task use different graph layouts to assess the models' ability to generalize across visual representations.

Number of graphs: more than 600K in total
File formats: '.png' and '.txt'

Each 'image_{i}.png' in the input directories corresponds to the i-th row of its associated 'label.txt' file.