Comparing Graph Convolutional Networks with Poisson Learning algorithm
Our research primarily consisted of applying Graph-based semi-supervised learning algorithms such as Laplace Learning, Poisson Learning and Graph Convolutional Networks to various datasets for multi-class classification and examining their efficacy. We used a variety of datasets such as Cora, CiteSeer, PubMed and MSTAR to name a few. We observed that at a fixed labeling rate, one algorithm worked better than other algorithms depending on the dataset. For instance, we find that Poisson Learning works better on PubMed than any other algorithm at low label rates, meanwhile for MSTAR, Laplace Learning seems to do better than Poisson Learning even at low label rates. This is the central question that our research was set out to examine and try to answer. In this UROP project, we look at various properties of the graphs themselves to determine what properties make the graph more suitable to a particular graph-based algorithm. Some of the properties we examine are diameter of the graph, the degree distribution histogram, the PageRank scores histogram, eigenvector centrality histogram and closeness centrality histograms.