CLP-GCN: Confidence and label propagation applied to Graph Convolutional Networks

Node classification is an important task in many graph-related applications. Recently, graph neural networks like Graph Convolutional Networks (GCNs) have provided low-dimensional informative representations of the nodes of the graphs. However, there are some issues with the application of GCNs in real-world problems. First, information sharing among nodes is limited to directly connected nodes at each layer of GCNs and the contribution of all neighbors of each node is the same in feature aggregation. Second, learning a new representation and training a classifier over this representation require a considerable amount of labeled nodes while there are few labeled nodes in many real world node classification tasks. To compensate for the lack of enough labeled data, in this paper, we propose a novel method for label propagation for GCNs. The main idea of this method, named CLP-GCN, is to propagate the confidence scores of the labels along with the pseudo-labels assigned to the nodes. Two closeness criteria are used to estimate the similarity of each pair of nodes and this similarity measure governs the label propagation from one node of the graph to its neighbors. The confidence of each pseudo-label is determined based on the confidences of the adjacent nodes with the same label. Moreover, the adjacency matrix of GCN is replaced with a more informative matrix representing the structural closeness of the nodes resulting in a more selective feature smoothing in this network. Comprehensive experiments on popular datasets like Citation, Co-authorship, Copurchase, and knowledge graph datasets demonstrate the superiority of CLP-GCN over state-of-the-art methods.