Abstract

Abstract

Networks are an effective way of representing complex data structures such as neural connectomes. When working with a large collection of networks, graph kernels measure how similar or dissimilar one graph is to another, which can help us with clustering disorganized data points into a lower dimensional structure that we can visualize and understand. There are many complex graph kernels that have been developed, but they can be computationally expensive when working with large datasets. Here we build ”simple” kernels that require much less computational power using properties of the network and provide an overview of their performances on weighted matched and unmatched networks, based on discriminability and different clustering methods. The results are then compared to well-established methods. For matched networks, we find that kernels based on the known matching of nodes perform better than kernels based on global properties of the network, and for unmatched metworks, we observe the opposite. This suggests that kernel performance may depend more on network properties than kernel complexity. More datasets should be explored to solidfy any identified trends, but these findings suggest that simple kernels can be a much faster yet still relatively effective way of measuring dissimilarity between graphs.