How Powerful are Graph Neural Networks? — mentioned by Topolomancer

Mon, 06 Jul 2020 20:00:15 GMT

Sure thing! For the GIN-ε (https://arxiv.org/pdf/1810.00826.pdf), for example, the authors report a classification accuracy of 75.9±3.8 on the PROTEINS data set (classical graph benchmark data set). If you run it with a cross-validation setup that is repeated to account for effects of chance, performance drops to 73.1±0.7. Notice the drop—the second accuracy value is at least within the standard deviation of the first one, but you can see that a different experimental setup shrinks the gains quite a lot... Same goes for different data sets. Since the gains are not super large for most papers…

Paper and discussion history. A mention is not an endorsement.

How Powerful are Graph Neural Networks? · HackerPapers discussions

How Powerful are Graph Neural Networks? — mentioned by Topolomancer