Bayesian graph convolutional neural networks

1. A computer implemented method for predicting labels for nodes in an observed graph, comprising:

deriving a plurality of random graph realizations of the observed graph;

learning a predictive function using the random graph realizations;

predicting label probabilities for nodes of the random graph realizations using the learned predictive function;

averaging the predicted label probabilities to predict labels for the nodes of the observed graph.

2. The method of claim 1 wherein deriving a set of random graph realizations comprises:

learning an generative graph function based on the observed graph, the generative graph function being configured to generate a plurality of probability matrices that each include a respective set of probability values for connections between nodes of the observed graph;

sampling the observed graph using the plurality of probability matrices to generate a respective set of random graph realizations corresponding to each of the probability matrices.

3. The method of claim 2 wherein the generative graph function comprises an assortative mixed membership stochastic block model (a-MMSBM).

4. The method of claim 3 wherein the sampling is Bernoulli sampling.

5. The method of claim 1 wherein:

learning the predictive function using the random graph realizations comprises learning, for each of the random graph realizations, a respective set of function parameters for the predictive function; and

predicting label probabilities for nodes of the random graph realizations comprises predicting, for each random graph realization, the label probabilities for the nodes using the respective set of function parameters learned for the random graph realization.

6. The method of claim 5 wherein the predictive function is learned using a graph convolution neural network (GCNN) and the function parameters include weights applied at convolution neural network layers of the GCNN.

7. The method of claim 6 wherein the respective set of function parameters for each random graph realization includes multiple sets of weights learned for the predictive function, wherein the label probabilities predicted for the nodes of each random graph realization includes a plurality of probabilities predicted based on each of the multiple sets of weights.

8. The method of claim 7 wherein the multiple sets of weights learned in respect of each random graph realization model are derived from a common set of weights using a Monte Carlo dropout.

9. The method of claim 6 wherein the predictive function is configured to perform a classification task and the labels predicted for the nodes specify a class from a plurality of possible classes.

10. The method of claim 6 wherein the predictive function is configured to perform a regression task and the labels predicted for the nodes specify a real-valued response variable.

11. The method of claim 1 wherein the observed graph is represented as an observed node feature matrix that includes feature vectors in respect of each of the nodes and an observed adjacency matrix that defines connections between the nodes, a subset of the nodes having labels, wherein deriving a plurality of random graph realizations of the observed graph comprises generating a plurality of constrained random variations of the observed node feature matrix.

12. A processing unit for predicting labels for nodes in an observed graph, the processing unit comprising a processing device and a storage storing instructions for configuring the processing unit to:

derive a plurality of random graph realizations of the observed graph;

learn a predictive function using the random graph realizations;

predict label probabilities for nodes of the random graph realizations using the learned predictive function;

average the predicted label probabilities to predict labels for the nodes of the observed graph.

13. The processing unit of claim 12 wherein the instructions configure the processing unit to derive a set of random graph realizations by:

learning an generative graph function based on the observed graph, the generative graph function being configured to generate a plurality of probability matrices that each include a respective set of probability values for connections between nodes of the observed graph;

sampling the observed graph using the plurality of probability matrices to generate a respective set of random graph realizations corresponding to each of the probability matrices.

14. The processing unit of claim 13 wherein the generative graph function comprises an assortative mixed membership stochastic block model (a-MMSBM).

15. The processing unit of claim 12 wherein:

the instructions configure the processing unit to learn the predictive function using the random graph realizations by learning, for each of the random graph realizations, a respective set of function parameters for the predictive function; and

the instructions configure the processing unit to predict the label probabilities for nodes of the random graph realizations by predicting, for each random graph realization, the label probabilities for the nodes using the respective set of function parameters learned for the random graph realization.

16. The processing unit of claim 15 wherein the predictive function is learned using a graph convolution neural network (GCNN) and the function parameters includes weights applied at convolution neural network layers of the GCNN.

17. The processing unit of claim 16 wherein the respective set of function parameters for each random graph realization includes multiple sets of weights learned for the predictive function, wherein the label probabilities predicted for the nodes of each random graph realization includes a plurality of probabilities predicted based on each of the multiple sets of weights.

18. The processing unit of claim 16 wherein the predictive function is configured to perform a classification task and the labels predicted for the nodes specify a class from a plurality of possible classes.

19. The processing unit of claim 16 wherein the predictive function is configured to perform a regression task and the labels predicted for the nodes specify a real-valued response variable.

20. A machine learning system comprising:

a graph generation module configured to receive as inputs an observed graph and output a plurality of random graph realizations of the observed graph;

a graph convolution neural network configured to learn a predictive function using the random graph realizations to predict label probabilities for nodes of the random graph realizations, and average the predicted label probabilities to predict labels for the nodes of the observed graph.