PyG-SSL Algorithms

Methods

Please refer to the “Tutorial” section and our provided jupiter notebooks for the detailed usage of the algorithms.

Neural Networks

MLPnum_layers, input_dim, hidden_dim, output_dim, activation="relu", norm="batchnorm")

Multi-layer perceptron (MLP) model.

Parameters:

• num_layers (int):

  The number of layers.

• input_dim (int):

  The input feature dimension.

• hidden_dim (int):

  The hidden feature dimension.

• output_dim (int):

  The output feature dimension.

• activation (str):

  The activation function. Default is “relu”.

• norm (str):

  The type of normalization layer. Default is “batchnorm”.

class NormLayer(hidden_dim, norm_type)

Normalization layer for the neural networks.

Parameters:

• hidden_dim (int):

  The hidden dimension of the normalization layer.

• norm_type (str):

  The type of normalization layer. Default is “batch”.

class ApplyNodeFunc(mlp, norm='batchnorm', activation='relu')

Apply a function to the node features.

Parameters:

• mlp (torch.nn.Module):

  The multi-layer perceptron.

• norm (str):

  The type of normalization layer. Default is “batchnorm”.

• activation (str):

  The activation function. Default is “relu”.

class GAT(in_dim, num_hidden, out_dim, num_layers, nhead, nhead_out, activation, feat_drop, attn_drop, negative_slope, residual, norm, concat_out=False, encoding=False)

Graph Attention Networks (GAT) from the “Graph Attention Networks” paper.

Parameters:

• in_dim (int):

  The input feature dimension.

• num_hidden (int):

  The hidden feature dimension.

• out_dim (int):

  The output feature dimension.

• num_layers (int):

  The number of layers.

• nhead (int):

  The number of head attentions in layer.

• nhead_out (int):

  The number of head attentions in the output layer.

• activation (Callable):

  The activation function.

• feat_drop (float):

  The dropout rate for feature.

• attn_drop (float):

  The dropout rate for attention.

• negative_slope (float):

  The negative slope of leaky ReLU.

• residual (bool):

  Use residual connection or not.

• norm (str):

  The type of normalization layer.

• concat_out (bool):

  Concatenate the output of all heads or not.

• encoding (bool):

  Use encoding mode or not.

class DotGAT(in_dim, num_hidden, out_dim, num_layers, nhead, nhead_out, activation, feat_drop, attn_drop, residual, norm, concat_out=False, encoding=False)

The GAT model with dot-product attention.

Parameters:

• in_dim (int):

  The input feature dimension.

• num_hidden (int):

  The hidden feature dimension.

• out_dim (int):

  The output feature dimension.

• num_layers (int):

  The number of layers.

• nhead (int):

  The number of head attentions in layer.

• nhead_out (int):

  The number of head attentions in the output layer.

• activation (Callable):

  The activation function.

• feat_drop (float):

  The dropout rate for feature.

• attn_drop (float):

  The dropout rate for attention.

• residual (bool):

  Use residual connection or not.

• norm (str):

  The type of normalization layer.

• concat_out (bool):

  Concatenate the output of all heads or not.

• encoding (bool):

  Use encoding mode or not.

class GCN(in_dim, num_hidden, out_dim, num_layers, dropout, activation, residual, norm, encoding=False)

The Graph Convolutional Networks (GCN) model.

Parameters:

• in_dim (int):

  The input feature dimension.

• num_hidden (int):

  The hidden feature dimension.

• out_dim (int):

  The output feature dimension.

• num_layers (int):

  The number of layers.

• dropout (float):

  The dropout rate.

• activation (Callable):

  The activation function.

• residual (bool):

  Use residual connection or not.

• norm (str):

  The type of normalization layer.

• encoding (bool):

  Use encoding mode or not.

class GIN(in_dim, num_hidden, out_dim, num_layers, dropout, activation, residual, norm, encoding=False, learn_eps=False, aggr='sum')

The Graph Isomorphism Network (GIN) model.

Parameters:

• in_dim (int):

  The input feature dimension.

• num_hidden (int):

  The hidden feature dimension.

• out_dim (int):

  The output feature dimension.

• num_layers (int):

  The number of layers.

• dropout (float):

  The dropout rate.

• activation (Callable):

  The activation function.

• residual (bool):

  Use residual connection or not.

• norm (str):

  The type of normalization layer.

• encoding (bool):

  Use encoding mode or not.

• learn_eps (bool):

  Learn the epsilon parameter or not.

• aggr (str):

  The type of aggregation function.