ScalarGCN: scalar-value association analysis of volumes based on graph convolutional network.

The relationships in multivariable data are intricate, and there are usually implicit associations between scalar values variables. However, existing association analysis methods lack spatial measurement of scalar values, and fail to collaboratively analyze the association between scalar values and variables. Thus association results may be one-sided. In this paper, we construct a scalar-value neighborhood graph to preserve the spatial information for scalar values and propose a graph neural network model composed of multiple graph convolutional layers and a self-attention mechanism for learning the low-dimensional vectors of scalar values and variables simultaneously. Several case studies show the scalability and flexibility of our method on analyzing the association between scalar values and variables. [ABSTRACT FROM AUTHOR]