“…The standard approach for assessing the generalization performance of a neural spatial interaction model is data splitting (see, for example, Fischer and Reismann, 2002b). This method simulates learning and generalization by partitioning the total data set, say M U ={(x u , y u ) with u=1, ..., U}, into three separate subsets: a training [insample] set M U1 ={(x u1 , y u1 ) with u1=1, ..., U1}, an internal validation set M U2 ={(x u2 , y u2 ) with u2=1, ..., U2} and a testing [out-of-sample] set M U3 ={(x u3 , y u3 ) with u3=1, ..., U3}.…”