Semi-Supervised Learning Without a Graph. Our robustness experiments show that GAM is effective even when the majority of edges in the graph connect nodes with mismatched labels. Therefore, we tested its power further by considering a more extreme scenario: no graph is provided, and the agreement model is tasked with learning whether an arbitrary pair of nodes shares a label. Note that having no graph, and picking random pairs of samples to use in the regularization terms in Equation 1, is equivalent to having a fully-connected graph from which we sample edges. We tested this scenario on Xxxx, Citeseer, and Pubmed and the results are marked as GAM* in Table 1. For completeness, we also show results for GCN+GAM* and GAT+GAM*, where even though the GAM* regularization term does not use the graph, the classification models use it by design. Our results show that GAM* also boosts the performance of all tested baseline models, with a gain of up to 19% accuracy for MLPs, 3.3% for GCNs, and 4.6% for GATs. It is worth noting that, even though GAM outperforms GAM* due to the extra information provided by the graph, GAM* generally outperforms the competing methods that also do not use a graph, and often even NGM which does. Non-graph Datasets. Since our approach no longer requires a graph to be provided, we tested GAM on the popular CIFAR-10 [16] and SVHN [26] datasets. For evaluation, we use the setup and train/validation/test splits provided by [27], which aims to provide a realistic framework for evaluating SSL methods. Thus, we start with 4000 and 1000 labeled samples for CIFAR-10 and SVHN, respectively, while the remaining training samples are considered unlabeled. More information about these datasets can be found in Appendix B. It is important to note that while Xxxx, Citeseer and Pubmed were evaluated under a transductive setting (where the input features and the graph structure of the test nodes are seen during training, but not their labels) as is typical in graph-based SSL, in the following experiments we evaluate GAM* under an inductive setting (the features of the test nodes are completely held out, and there is no graph to provide other information about them).
