Fig. 15
From: Extended study on atomic featurization in graph neural networks for molecular property prediction
Distributions of feature importance scores for three GCN models trained on Rat dataset with representation F-C which differ only by weight initialization (run-1 - run-3). For each feature type, the mean of the corresponding distributions and the standard deviation is shown on the right (averaged). Models which differ only by initial weight values may show differences in their exploitation of information about inclusion in a ring and aromaticity while for other features the variability between models is lower