Exploiting vertices states in GraphESN by weighted nearest neighbor

Graph Echo State Networks (GraphESN) extend the Reservoir Computing approach to directly process graph structures. The reservoir is applied to every vertex of an input graph, realizing a contractive encoding process and resulting in a structured state isomorphic to the input. Whenever an unstructured output is required, a state mapping function maps the structured state into a fixed-size feature representation that feeds the linear readout. In this paper we propose an alternative approach, based on distance-weighted nearest neighbor, to realize a more flexible readout exploiting the state information computed for every vertex according to its individual relevance.