We consider a family of problems that are concerned about making predictions for the majority of unlabeled, graph-structured data samples based on a small proportion of labeled samples. Relational information among the data samples, often encoded in the graph/network structure, is shown to be helpful for these semi-supervised learning tasks. However, conventional graph-based regularization methods and recent graph neural networks do not fully leverage the interrelations between the features, the graph, and the labels. In this work, we propose a flexible generative framework for graph-based semi-supervised learning, which approaches the joint distribution of the node features, labels, and the graph structure. Borrowing insights from random graph models in network science literature, this joint distribution can be instantiated using various distribution families. For the inference of missing labels, we exploit recent advances of scalable variational inference techniques to approximate the Bayesian posterior. We conduct thorough experiments on benchmark datasets for graph-based semi-supervised learning. Results show that the proposed methods outperform state-of-the-art models under most settings.

This work employs techniques developed in network science literature, such as latent space model (LSM) and stochastic block model (SBM), to propose a generative model for features X, outputs Y, and graph G, and it uses graph neural networks to approximate the posterior of missing outputs given X, observed Y, and G. This work is a wise combination of recent methods to effectively address the problem of graph-based semi-supervised learning. However, I have some concerns, which are summarized as follows: - Although the paper proposed a new interesting generative method for graph-based semi-supervised learning, it is not super novel, as it employs the other existing methods as the blocks of their method, like LSM, SBM, GCN, GAT. - It seems the generative model is only generative for G given X and Y and by factorizing the other part as p(Y,X) p(Y X) p(X), for p(Y X), it is modeled via a multi-layer perceptron, which is a discriminative model. That is why the authors discard X in all the analyses, like any other discriminative model, and say that everything is conditioned on X. I think this makes the proposed model not fully generative. It is only generative for G but not for X and Y.

This paper proposes a generative framework for graph-based semi-supervised learning for approximating the joint distribution of the graph structure, labels and the node features. Variational inference techniques are then used to approximate the Bayesian posterior. The paper is well written. There are some issues raised by reviewer 3 regarding a better positioning of GenGNN with respect to GCN/GAT; which are recommended to be taken into account for the final version of the paper.

We consider a family of problems that are concerned about making predictions for the majority of unlabeled, graph-structured data samples based on a small proportion of labeled samples. Relational information among the data samples, often encoded in the graph/network structure, is shown to be helpful for these semi-supervised learning tasks. However, conventional graph-based regularization methods and recent graph neural networks do not fully leverage the interrelations between the features, the graph, and the labels. In this work, we propose a flexible generative framework for graph-based semi-supervised learning, which approaches the joint distribution of the node features, labels, and the graph structure. Borrowing insights from random graph models in network science literature, this joint distribution can be instantiated using various distribution families.

We consider a family of problems that are concerned about making predictions for the majority of unlabeled, graph-structured data samples based on a small proportion of labeled samples. Relational information among the data samples, often encoded in the graph/network structure, is shown to be helpful for these semi-supervised learning tasks. However, conventional graph-based regularization methods and recent graph neural networks do not fully leverage the interrelations between the features, the graph, and the labels. In this work, we propose a flexible generative framework for graph-based semi-supervised learning, which approaches the joint distribution of the node features, labels, and the graph structure. Borrowing insights from random graph models in network science literature, this joint distribution can be instantiated using various distribution families.