We study the generalisation properties of majority voting on finite ensembles of classifiers, proving margin-based generalisation bounds via the PAC-Bayes theory. These provide state-of-the-art guarantees on a number of classification tasks. Our central results leverage the Dirichlet posteriors studied recently by Zantedeschi et al. (2021) for training voting classifiers; in contrast to that work our bounds apply to non-randomised votes via the use of margins. Our contributions add perspective to the debate on the "margins theory" proposed by Schapire et al. (1998) for the generalisation of ensemble classifiers.

Traditional solutions prioritizing easy or hard samples lack the flexibility to handle such a variety simultaneously. Recent work has proposed a more reasonable data selection principle by examining the data's impact on the model's generalization loss.

Now we look into the term[(A+I)X]TV,:, which is the aggregated feature vectors within neighborhood N1 for nodes in the training set. Note that [(A + I)X]TS,: is a circulant matrix, therefore its inverse exists. Now consider an arbitrary training datapoint(v,yv) TV, and a perturbation added to the neighborhood N(v) of node v, such that the number of nodes with a randomly selected class labelyp Y 6=yv isδ1lessthanexpectedin N(v). Now we move on to discuss the GCN layer formulated asf(X;A,W) = AXW without self loops. We regardcs,i as the coefficient ofs at frequency componenti and regard the coefficients at all frequencies components{cs,i} as the spectrum of signalswith respect to graphG.

For a small training set, CNNs behave similarly.