Fig. 6
From: Regularized Simple Graph Convolution (SGC) for improved interpretability of large datasets
The application of the regularized SGC to the dataset where linear projections are incapable of class separation. a Shows the parameter vectors produced by introducing constraints into SGC method on the feature space with original data points and how the classification can then produce perfect accuracy. It can be seen the proposed SGC reduces the effective number of features used in the columns of the matrix \(\varvec{\Theta }_R\). b Shows the plots for the same set of weight vectors displayed on the projection axes of \({\mathbf {S}}^2 {\mathbf {X}}\) where each datapoint (node) accumulates feature information from neighbors 2-hops away (\(k=2\))