Most graph neural network architectures take the input graph as granted and do not assign any uncertainty to its structure. In real life, however, data is often noisy and may contain incorrect edges or exclude true edges. Bayesian methods, which consider the input graph as a sample from a distribution, have not been deeply researched, and most existing research only tests the methods on small benchmark datasets such as citation graphs. As often is the case with Bayesian methods, they do not scale well for large datasets.
The goal of this thesis is to research different Bayesian graph neural network architectures for semi-supervised node classification and test them on larger datasets, trying to find a method that improves the baseline model and is scalable enough to be used with graphs of tens of thousands of nodes with acceptable latency. All the tests are done twice with different amounts of training data, since Bayesian methods often excel with low amounts of data and in real life labeled data can be scarce. The Bayesian models considered are based on the graph convolutional network, which is also used as the baseline model for comparison.
This thesis finds that the impressive performance of the Bayesian graph neural networks does not generalize to all datasets, and that the existing research relies too much on the same small benchmark graphs. Still, the models may be beneficial in some cases, and some of them are quite scalable and could be used even with moderately large graphs.
