Graph Neural Networks

Graph Neural Networks are a class of neural networks, optimized for deep learning on graphs. They provide an end-to-end solution for machine learning on graphs, unlike graph kernels/embeddings where a transformation step is applied before a "classical" machine learning algorithm. As they have attracted more attention in the recent years, a range of different architectures for this class has sprung up ranging from the Graph Convolutional Network (GCN, Kipf et al. 2017) to Graph Attention Networks (citation needed). The different architectures learn from the graph and it's overall structure, making use of the graph information unlike classical neural networks.

The graph neural network module of PHOTONAI Graph provides a variety of customizable out-of-the-box graph neural networks. They can be instantiated in one line of code and easily integrate into PHOTONAI pipelines.

Graph Neural Network Module

The Graph Neural Network module consists of three parts. The Layer Module, where different layers are implemented and the message-passing steps of these are defined. The Model module where the module is constructed as a class (see pytorch neural networks). And the GraphConvNet module which calls the models and implements fit and transform steps, making them sklearn conform. This module also handles data conversions, converting graphs to the right format for the networks, which are written in pytorch.

You can also write your own custom graph neural network architecture, and register them via the PHOTON register function (link here). When writing your own custom neural nets you are free to choose your own package, as long as they implement fit, transform and predict functions like the GraphConvNet module classes. These can also be used as a blueprint if you want to integrate your own graph neural network architectures into PHOTONAI.

DglModel

Base class for DGL based graph neural networks. Implements helper functions and shared parameters used by other models. Implementation based on dgl python package.

Parameters:

Name	Type	Description	Default
nn_epochs	int	the number of epochs which a model is trained	200
learning_rate	float	the learning rate when training the model	0.001
batch_size	int	number of samples per training batch	32
adjacency_axis	int	position of the adjacency matrix, default being zero	0
feature_axis	int	position of the feature matrix	1
add_self_loops	bool	self loops are added if true	True
allow_zero_in_degree	bool	If true the zero in degree test of dgl is disabled	False
validation_score	bool	If true the input data is split into train and test (90%/10%). The testset is then used to get validation results during training	False
early_stopping	bool	If true then the loss over multiple iterations is evaluated to see whether early stopping should be called on the model	False
verbose	bool	If true verbose information is printed	False
logs	str	Path to the log data	None

DGLRegressorBaseModel

Abstract base class for regression algorithms

Parameters:

Name	Type	Description	Default
nn_epochs	int	Number of epochs to fit the model	200
learning_rate	float	Learning rate for model training	0.001
batch_size	int	Batch size for model training	32
adjacency_axis	int	Axis which contains the adjacency	0
feature_axis	int	Axis which contains the features	1
add_self_loops	bool	If this value is true, a self loop is added to each node of each graph	True
allow_zero_in_degree	bool	If true the dgl model allows zero-in-degree Graphs	False
validation_score	bool	If true the input data is split into train and test (90%/10%). The testset is then used to get validation results during training	False
verbose	bool	If true verbose output is generated	False
logs	str	Default logging directory	None

DGLClassifierBaseModel

Abstract base class for classification algorithms

Parameters:

Name	Type	Description	Default
nn_epochs	int	Number of epochs to fit the model	200
learning_rate	float	Learning rate for model training	0.001
batch_size	int	Batch size for model training	32
adjacency_axis	int	Axis which contains the adjacency	0
feature_axis	int	Axis which contains the features	1
add_self_loops	bool	If this value is true, a self loop is added to each node of each graph	True
allow_zero_in_degree	bool	If true the dgl model allows zero-in-degree Graphs	False
validation_score	bool	It true the input data is split into train and test (90%/10%). The testset is then used to get validation results during training	False
verbose	bool	If true verbose output is generated	False
logs	str	Default logging directory	None

GCNClassifierModel

Graph Attention Network for graph classification. GCN Layers from Kipf & Welling, 2017. Implementation based on dgl & pytorch.

Parameters:

Name	Type	Description	Default
in_dim	int	input dimension	1
hidden_layers	int	number of hidden layers used by the model	2
hidden_dim	int	dimensions in the hidden layers	256
validation_score	bool	If true the input data is split into train and test (90%/10%). The testset is then used to get validation results during training	False
verbose	bool	If true verbose output is generated	False

SGConvClassifierModel

Graph convolutional network for graph classification. Simple Graph convolutional layers from Wu, Felix, et al., 2018. Implementation based on dgl & pytorch.

Parameters:

Name	Type	Description	Default
in_dim	int	input dimension	1
hidden_layers	int	number of hidden layers used by the model	2
hidden_dim	int	dimensions in the hidden layers	256
validation_score	bool	If true the input data is split into train and test (90%/10%). The testset is then used to get validation results during training	False
verbose	bool	If true verbose output is generated	False

GATClassifierModel

Graph Attention Network for graph classification. GAT Layers are modeled after Veličković et al., 2018. Implementation based on dgl & pytorch.

Parameters:

Name	Type	Description	Default
in_dim	int	input dimension	1
hidden_layers	int	number of hidden layers used by the model	2
hidden_dim	int	dimensions in the hidden layers	256
heads	List	list with number of heads per hidden layer	None
validation_score	bool	If true the input data is split into train and test (90%/10%). The testset is then used to get validation results during training	False
verbose	bool	If true verbose output is generated	False
agg_mode		aggregation mode for the graph convolutional layers	'mean'

GCNRegressorModel

Graph convolutional Network for graph regression. GCN Layers from Kipf & Welling, 2017. Implementation based on dgl & pytorch.

Parameters:

Name	Type	Description	Default
in_dim	int	input dimension	1
hidden_layers	int	number of hidden layers used by the model	2
hidden_dim	int	dimensions in the hidden layers	256
validation_score	bool	If true the input data is split into train and test (90%/10%). The testset is then used to get validation results during training	False
verbose	bool	If true verbose output is generated	False

SGConvRegressorModel

Graph convolutional network for graph regression. Simple Graph convolutional layers from Wu, Felix, et al., 2018. Implementation based on dgl & pytorch.

Parameters:

Name	Type	Description	Default
in_dim	int	input dimension	1
hidden_layers	int	number of hidden layers used by the model	2
hidden_dim	int	dimensions in the hidden layers	256
validation_score	bool	If true the input data is split into train and test (90%/10%). The testset is then used to get validation results during training	False
verbose	bool	If true verbose output is generated	False

GATRegressorModel

Graph Attention Network for graph regression. GAT Layers are modeled after Veličković et al., 2018. Implementation based on dgl & pytorch.

Parameters:

Name	Type	Description	Default
in_dim	int	input dimension	1
hidden_layers	int	number of hidden layers used by the model	2
hidden_dim	int	dimensions in the hidden layers	256
heads	List	list with number of heads per hidden layer	None
validation_score	bool	If true the input data is split into train and test (90%/10%). The testset is then used to get validation results during training	False
verbose	bool	If true verbose output is generated	False