Multi-Stage Self-Supervised Learning for Graph Convolutional Networks on Graphs with Few Labeled Nodes

Abstract

Graph Convolutional Networks (GCNs) play a crucial role in graph learning tasks, however, learning graph embedding with few supervised signals is still a difficult problem. In this paper, we propose a novel training algorithm for Graph Convolutional Network, called Multi-Stage Self-Supervised (M3S) Training Algorithm, combined with self-supervised learning approach, focusing on improving the generalization performance of GCNs on graphs with few labeled nodes. Firstly, a Multi-Stage Training Framework is provided as the basis of M3S training method. Then we leverage DeepCluster technique, a popular form of self-supervised learning, and design corresponding aligning mechanism on the embedding space to refine the Multi-Stage Training Framework, resulting in M3S Training Algorithm. Finally, extensive experimental results verify the superior performance of our algorithm on graphs with few labeled nodes under different label rates compared with other state-of-the-art approaches.

Investigating and Mitigating Degree-Related Biases in Graph Convoltuional Networks

“…6.1.1 Datasets. For a fair comparison, we adopt same benchmark datasets used by Sun et al [24] and Li et al [17], including Cora, Citeseer, Pubmed [22]. Each dataset contains a citation graph, where nodes represent articles/papers and edges denote citation correlation.…”

“…For example, DCNN [2] combines graph convolutional operator with the diffusion process and Veličković et al proposes the graph attention network [28] with the self-attention mechanism on the neighbors of the node and assign different weights accordingly during the aggregation process. Of all these GCNNs, GNNs [15] are highly favorable by the computer science community [17,24]…”

“…Though still in its early stage, there are a few seminal work trying to exploit self-supervised training to improve the performance of GCNs [12]. For example, Li et al [17] propose the co-training and self-training based GCN models by expanding the training node set with pseudo labels from its nearest neighborhoods; Sun et al [24] combine DeepCluster [5] with a multi-stage training framework so that the generalization performance of GCNs with few labeled nodes are improved.…”

“…The intuition of proposed self-learning algorithm is to leverage the large amount of pseudo-labels in the training of GCNs so that even for low-degree nodes can have labeled neighbors. However, different from existing literature [17,24] that use pseudo labels in the same way of true labeled nodes, we also judge the quality of pseudo labels to avoid overfitting on inaccurate pseudo labels. Specifically, we design a Bayesian neural network as a teacher to justify the quality of pseudo-labels from the annotator, so that the GCNs as a student can fully exploit the pseudo-labels.…”

“…Existing works exploring self-learning for GNNs treat selected pseudo labels in the same way of using labeled nodes. For example, Li et al [17] and Sun et al [24] progressively add selected nodes with pseudo labels into the training set. However, such solutions are sub-optimal.…”