A Methodology for Neural Spatial Interaction Modeling

摘要: This paper attempts to develop a mathematically rigid and unified framework for neural spatial interaction modeling. Families of classical neural network models, but also less ones such as product unit are considered for the cases unconstrained singly constrained flows. Current practice appears suffer from least squares normality assumptions that ignore the true integer nature the flows approximate discrete-valued process by an almost certainly misrepresentative continuous distribution. To overcome this deficiency we suggest more suitable estimation approach, maximum likelihood under more realistic distributional Poisson processes, utilize global search procedure, called Alopex, solve problem. To identify transition underfitting overfitting we split data into training, internal validation test sets. The bootstrapping pairs approach with replacement is adopted combine purity splitting power resampling procedure generally neglected issue fixed the problem scarce data. In addition, has provide better statistical picture prediction variability, Finally, benchmark comparison against gravity models illustrates superiority both, the unconstrained origin model versions in terms of generalization performance measured Kullback Leibler's information criterion.