2019/09/17 by Sergio Garrido, Garrido, Sergio, Stanislav S. Borysov +5 · 5 citations
Engineering · Social Sciences · #Applications (stat.AP) #FOS: Computer and information sciences #Machine Learning (cs.LG) #Machine Learning (stat.ML) #Transportation Planning and Optimization #Transportation and Mobility Innovations #Urban Transport and Accessibility
paper · pdf · doi:10.48550/arxiv.1909.07689
openalex publication_date 2019/09/17 · openalex created_date 2022/07/28 · openalex updated_date 2026/07/28
In population synthesis applications, when considering populations with many\nattributes, a fundamental problem is the estimation of rare combinations of\nfeature attributes. Unsurprisingly, it is notably more difficult to reliably\nrepresentthe sparser regions of such multivariate distributions and in\nparticular combinations of attributes which are absent from the original\nsample. In the literature this is commonly known as sampling zeros for which no\nsystematic solution has been proposed so far. In this paper, two machine\nlearning algorithms, from the family of deep generative models,are proposed for\nthe problem of population synthesis and with particular attention to the\nproblem of sampling zeros. Specifically, we introduce the Wasserstein\nGenerative Adversarial Network (WGAN) and the Variational Autoencoder(VAE), and\nadapt these algorithms for a large-scale population synthesis application. The\nmodels are implemented on a Danish travel survey with a feature-space of more\nthan 60 variables. The models are validated in a cross-validation scheme and a\nset of new metrics for the evaluation of the sampling-zero problem is proposed.\nResults show how these models are able to recover sampling zeros while keeping\nthe estimation of truly impossible combinations, the structural zeros, at a\ncomparatively low level. Particularly, for a low dimensional experiment, the\nVAE, the marginal sampler and the fully random sampler generate 5%, 21% and\n26%, respectively, more structural zeros per sampling zero generated by the\nWGAN, while for a high dimensional case, these figures escalate to 44%, 2217%\nand 170440%, respectively. This research directly supports the development of\nagent-based systems and in particular cases where detailed socio-economic or\ngeographical representations are required.\n