Exploring Transit Use during COVID-19 Based on XGB and SHAP Using Smart Card Data

Abstract

As the coronavirus (COVID-19) pandemic continues, many protective measures have been taken in Seoul, Korea, and around the world. This situation has drastically changed lifestyle and travel behavior. An important issue concerns understanding the reasons for giving up transit use and the potential impact on travel patterns during the COVID-19 pandemic. To shed light on these issues that are essential for transit policy, this study explores transit use choice, such as whether users have given-up transit use or not, during the COVID-19 pandemic. Two days of smart card data, before and during the COVID-19 pandemic, were used to look at users who gave up transit use during the COVID-19 pandemic. The choice set of the dataset includes two alternatives, for example, transit use and given-up transit use. An extreme gradient boosting (XGB) model was used to estimate the transit use behavior. Shapley additive explanations were performed to interpret the estimation results of the XGB model. The results for the overall specificity, sensitivity, and balanced accuracy of the proposed XGB model were estimated to be 0.909, 0.953, and 0.931, respectively. The feature analysis based on the Shapley value shows that the number of origin-to-destination trip feature substantially impacts transit use. As such, users tend to avoid transit use as travel time increased during the COVID-19 pandemic. The proposed model shows remarkable performance in accuracy and provided an understanding of the estimated results.

Social and spatial heterogeneities in COVID-19 impacts on individual's metro use: A big-data driven causality inference

“…To summarize, although there has been increasing investigation of metro use adaptation to COVID-19 and mobility intervention policies, several research gaps that deserve further investigation remain. First, compared to studies that examine aggregate metro use response to COVID-19 at levels of metro stations, neighborhoods, and cities ( Carrión et al, 2021 ; Chang et al, 2021 ; Kwon et al, 2022 ; Mützel & Scheiner, 2022 ), few studies have observed the individual-level metro use responses, particularly using a large sample size ( Lee, 2022 ; Park & Cho, 2021 ). The regional or aggregate responses, which are often measured by the ridership change, are not equivalent to individuals' metro use behavioral reactions ( Sy et al, 2021 ).…”

“…Third, the limited attention given to the causality at the individual level might be due to the lack of longitudinal observations of individuals' repeated metro use, particularly right before and after the occurrence of the outbreak and mobility intervention events. Most existing studies rely on data from questionnaire surveys or interviews and transit smart card big data to investigate the associations between the COVID-19 and individual's metro use ( Lee, 2022 ; Maljaee & Sameni, 2022 ; Park et al, 2022 ; Zhou et al, 2021 ). However, these data sources are often difficult to be applied to build rigorous quasi-experimental designs.…”

“…By contrast, the smart card data typically provides a time series of repeated observations of daily metro use information, whereas this type of data is mostly not rich enough to support individual-based research, but aggregated at the station level or merely focused on single-faceted socioeconomic groups (e.g., Maljaee & Sameni, 2022 ; Park & Cho, 2021 ). To be more specific, such datasets have limited information of the individual's socioeconomic attributes for control and other travel behavior outcomes but only the metro use ( Lee, 2022 ).…”

“…Furthermore, most existing individual-level studies heavily rely on descriptive statistics or correlation analyses to measure the associations between the outbreak and metro use ( Lee, 2022 ; Park & Cho, 2021 ; Park et al, 2022 ), whereas none of these studies explore causal effects. Without a causal inference, the detected associations might be biased, as they could not disentangle the mixed effects of panic effects from virus, policy effects from interventions, and other confounders from weather and holidays ( Hu & Chen, 2021 ).…”

Deep Multimodal Learning for Traffic Speed Estimation Combining Dedicated Short-Range Communication and Vehicle Detection System Data

“…For example, models that use the autoregressive integrated moving average (ARIMA) perform well for normal conditions based on the stationary assumption of the time series data, but it is difficult for them to reflect the nonstationary traffic data, especially the nonlinear relationships between traffic variables (2). Non-parametric machine learning methods have been used extensively to overcome the nonlinear problems of statistical models (9,10). However, machine learning models, such as the support vector machine (SVM) and the k-nearest neighbor (KNN), require statistical assumptions to derive manually designed features (8).…”