Analysis of the Usage Patterns of Public Bikes in Seoul with Clustering and the Spatial Model

Abstract

The COVID-19 pandemic has significantly affected the daily routines and transportation behaviors. Public bikes have emerged as a popular means of transportation, offering a safe and socially distanced option for travel. The usage patterns of Seoul’s public bike ‘Ttareungi’ have undergone changes during the pandemic which leads to the need to analyze the usage patterns during this period to understand the unique dynamics of user behavior. This study analyzes the user data of Seoul’s public bike system during the COVID-19 pandemic period from July 2021 to December 2021 when the Level 4 social distancing measures were implemented in Seoul from July 2021. The analysis focuses specifically on the morning hours between 7 a.m. to 10 a.m. due to a prominent spike in activity during this time period. By delving into this specific timeframe during the COVID-19 period, the study aims to provide valuable insights into the usage patterns during the morning hours amidst the pandemic. The two-stage approach was proposed in this study. The first stage applied the GMM (Gaussian Mixture Model) clustering method to cluster the bike stations based on their usage and environmental variables such as the distance to the nearest bus and subway stations from the bike stations. In the second stage, the obtained clustering results are integrated into the modeling process. Specifically, the clustering results derived from the first stage are reflected by the cluster coefficients in the model. Each cluster is assigned different coefficients, reflecting the distinct characteristics of the clusters. Moreover, the model also considers the spatial effects by incorporating the spatial association between the bike stations. Four models were considered for comparison. The performance of the models was assessed using the metrics MAE (Mean Absolute Error), MSPE (Mean Squared Prediction Error), DIC (Deviance information criteria), and MPL (Marginal predictive likelihood). The model that incorporated the clustering effect and the model that combined both the clustering and spatial effects demonstrated the high performance. Both of the models produced accurate results.