Spatial transferability of pedestrian trip generation models
研究了澳大利亚和美国五个城市行人出行生成模型的空间可转移性，发现聚合模型在特定条件下可转移，而随机森林模型在估计步行出行次数和跨城市转移方面表现较好，负二项回归则擅长处理高变异性数据。
The availability and consistency of pedestrian travel data vary across different locations, often requiring the transfer of estimated models in the absence of comprehensive local data. However, the extent to which pedestrian demand models are spatially transferable is not well understood. This study explores the spatial transferability of both aggregate and disaggregate pedestrian trip generation models using data from the Household Travel Surveys of Sydney, Melbourne, and Brisbane, Australia and two cities in the United States, Seattle and Chicago. We estimate Negative Binomial regression, Bayesian regression, and Random Forest models as aggregate approaches, while for disaggregate individual walking trip generation, we estimate a Poisson zero-inflated model, a two-step Logit-Bayesian approach, and a two-step Random Forest model. Results suggest that aggregate models exhibit reasonable transferability under certain conditions, while disaggregate models show greater limitations. The study demonstrates that while Random Forest generally outperforms other models in estimating the number of walking trips and shows strong transferability between cities, Negative Binomial Regression is effective at handling data with high variability, often surpassing machine learning models. The results highlight that both traditional and machine learning approaches have distinct advantages depending on data characteristics and under some data conditions such as sample size, the distribution of variables, and the heterogeneity of input variables. The combined use of these models can effectively capture the behavior of walking trip generation at different scales and provide valuable insights for policymakers and urban planners at both city-wide and localized levels, especially in areas where data might be lacking.