Towards a Framework to Evaluate Generative Time Series Models for Mobility Data Features

Abstract

Understanding human mobility has implications for several areas, such as immigration, disease control, mobile networks performance, and urban planning. However, gathering and disseminating mobility data face challenges such as data collection, handling of missing information, and privacy protection. An alternative to tackle these problems consists of modeling raw data to generate synthetic data, preserving its characteristics while maintaining its privacy. Thus, we propose MobDeep, a unified framework to compare and evaluate generative models of time series based on mobility data features, which considers statistical and deep learning-based modeling. To achieve its goal, MobDeep receives as input statistical or Generative Adversarial Network-based models (GANs) and the raw mobility data, and outputs synthetic data and the metrics comparing the synthetic with the original data. In such way, MobDeep allows evaluating synthetic datasets through qualitative and quantitative metrics. As a proof-of-concept, MobDeep implements one classical statistical model (ARIMA) and three GANs models. To demonstrate MobDeep on distinct mobility scenarios, we considered an open dataset containing information about bicycle rentals in US cities and a private dataset containing information about a Brazilian metropolis's urban traffic. MobDeep allows observing how each model performs in specific scenarios, depending on the characteristics of the mobility data. Therefore, by using MobDeep researchers can evaluate their resulting models, improving the fidelity of the synthetic data regarding the original dataset.