shapiq_benchmark.setups.GraphSetup

class shapiq_benchmark.setups.GraphSetup(*, dataset='MUTAG', model='gcn', n_layers=2, hidden_size=32, n_epochs=100, learning_rate=0.01, x=0, class_index=None, baseline='zeros', test_size=0.2, random_state=42, normalize=True, batch_size=128)

Bases: Setup

A GraphGame of a graph neural network trained on a graph dataset.

The network has n_layers message passing layers of the model with ReLU, a mean over the nodes, and a linear readout of the class logits, so its game is local on the graph and GraphComputer computes its exact values with GraphSHAP-IQ. It is trained with Adam on the training split of the dataset (from load_graph_dataset(), stratified); the explained graph is taken from the test split.

Variables:

• dataset – The graph dataset. Defaults to "MUTAG".

• model – The message passing layers, "gcn" (default), "gin", or "gat".

• n_layers – The number of message passing layers, the radius of the receptive fields. Defaults to 2.

• hidden_size – The size of the node embeddings. Defaults to 32.

• n_epochs – The number of training epochs. Defaults to 100.

• learning_rate – The learning rate of Adam. Defaults to 0.01.

• x – The index of the explained graph in the test split. Defaults to 0.

• class_index – The explained class (None means class 1).

• baseline – The node features of absent nodes, "zeros" (default) or the "mean" node features of the training graphs.

• test_size – The fraction of the graphs in the test split. Defaults to 0.2.

• random_state – The seed of the split, the initialization, and the training.

• normalize – Whether to center the game. Defaults to True.

• batch_size – The number of masked graphs per forward pass (a runtime field).

Examples

>>> setup = GraphSetup(dataset="MUTAG", model="gin", x=3)
>>> game = setup.build()

build()

Load the graphs, train the network, and build the game.

Return type:

GraphGame

to_dict()

Return the setup as a JSON-serializable dictionary (the inverse of setup_from_dict()).

Return type:

dict[str, Any]

baseline = 'zeros'

batch_size = 128

class_index = None

dataset = 'MUTAG'

hidden_size = 32

property key

A stable identifier of the game this setup builds, used as its cache key.

It is a digest of the setup’s name, version, RECIPE_VERSION, and fields (runtime fields excluded), so it is the same across processes and machines.

learning_rate = 0.01

model = 'gcn'

n_epochs = 100

n_layers = 2

name = 'graph'

The registered name of the setup.

normalize = True

random_state = 42

test_size = 0.2

version = 1

The version of build(), part of key.

x = 0