2024/01/18 by Fischer, Manfred M., Reismann, Martin
paper · doi:10.57938/a425753c-dccb-482c-b494-53f1ebc882ba
This paper exposes problems of the commonly used technique of splitting the available <br/>data in neural spatial interaction modelling into training, validation, and test sets that <br/>are held fixed and warns about drawing too strong conclusions from such static splits. <br/>Using a bootstrapping procedure, we compare the uncertainty in the solution stemming <br/>from the data splitting with model specific uncertainties such as parameter <br/>initialization. Utilizing the Austrian interregional telecommunication traffic data and <br/>the differential evolution method for solving the parameter estimation task for a fixed <br/>topology of the network model [ i.e. J = 9] this paper illustrates that the variation due to <br/>different resamplings is significantly larger than the variation due to different parameter <br/>initializations. This result implies that it is important to not over-interpret a model, <br/>estimated on one specific static split of the data. (authors' abstract)