GenPR: Generative PageRank Framework for Semi-supervised Learning on Citation Graphs

Abstract

Nowadays, Semi-Supervised Learning (SSL) on citation graph data sets is a rapidly growing area of research. However, the recently proposed graph-based SSL algorithms use a default adjacency matrix with binary weights on edges (citations), that causes a loss of the nodes (papers) similarity information. In this work, therefore, we propose a framework focused on embedding PageRank SSL in a generative model. This framework allows one to do joint training of nodes latent space representation and label spreading through the reweighted adjacency matrix by node similarities in the latent space. We explain that a generative model can improve accuracy and reduce the number of iteration steps for PageRank SSL. Moreover, we show that our framework outperforms the best graph-based SSL algorithms on four public citation graph data sets and improves the interpretability of classification results.

Graph Diffusion & PCA Framework for Semi-supervised Learning

“…The recent advances in GB-SSL can be classified into the following rapidly growing directions: (1) classical linear graph diffusion algorithms which apply the graph structure for spreading the information of labelled nodes through it, such as Label Propagation (LP) [33], PageRank SSL (PRSSL) [1], or manifold regularization (ManiReg) [4]; and (2) graph-convolution based neural network algorithms. The latter category can be further seperated into (i) nonlinear graph diffusion algorithms, which apply convolution on the graph's adjacency matrix A with node features, such as Graph Convolution Network (GCN) [19], approximated Personalized graph neural network (APPNP) [20], Planetoid [32], or DeepWalk [23]; and (ii) graph convolution deep generative models, focusing on the application of nonlinear graph convolution algorithms with respect to the latent representation of nodes/edges: GenPR [18], Graphite [13].…”