Learning graph structure via graph convolutional networks

Highlights：

• We proposed a framework of graph convolutional neural network which can handle the data defined on irregular or non-Euclidean domains directly.

• A graph structure learning method has been introduced to enable our method to learn the relationship between two adjacent nodes and grasp the graph structure information.

• The proposed method allows the convolution kernel weights to be tuned at different locations, which alleviates the restriction of weight sharing inherited from classical CNN and will be more suitable for the data without statistical stationarity.

• The non-linear activation function ReLU and the sparse constraint are employed on the graph structure parameters to promote the proposed method to focus on the important nodes in the neighborhoods and filter out the insignificant nodes.

• Extensive experiments have been explored at multiple datasets show the effectiveness of our method.

摘要

•We proposed a framework of graph convolutional neural network which can handle the data defined on irregular or non-Euclidean domains directly.•A graph structure learning method has been introduced to enable our method to learn the relationship between two adjacent nodes and grasp the graph structure information.•The proposed method allows the convolution kernel weights to be tuned at different locations, which alleviates the restriction of weight sharing inherited from classical CNN and will be more suitable for the data without statistical stationarity.•The non-linear activation function ReLU and the sparse constraint are employed on the graph structure parameters to promote the proposed method to focus on the important nodes in the neighborhoods and filter out the insignificant nodes.•Extensive experiments have been explored at multiple datasets show the effectiveness of our method.