Machine preparation for human labelling of hierarchical train sets by spectral clustering
Dávid Papp, Gábor Szücs, Zsolt Knoll · 2019
Human labeling of an unknown dataset for machine learning is a tedious work for humans. The aim of the paper was to develop a machine preparation of the data that helps human annotators to label at the higher level (point-set level). We present two different approaches to cluster point-sets with spectral clustering. The fundamental idea was to use the set of relations to rescale the weight of edges between point pairs. The first approach is based on a fully connected weight graph (FC-WG). In this case each point is connected to every other point, and only the weights control the outcome of the clustering. Another approach uses a proposed graph, so called the nearest points of point-sets weight graph (NPP-WG), which is not a fully connected graph, because the connections between point-sets are restricted. We investigated our theoretical approaches at two different datasets. The results show that spectral clustering with weight graphs are suitable to use on points, while it also provides the grouping of point-sets.