COVID-19 confines recreational gatherings in Seoul to familiar, less crowded, and neighboring urban areas

Human urban activities concentrate in specific city areas, forming a hierarchical structure where higher-level regions attract more visitors than lower-level ones. Such hierarchies raise concerns during pandemics: outbreaks from top hotspots can lead to widespread transmission, and economic impacts may differ across hierarchy levels, disproportionately affecting populations in lower tiers. Despite its relevance, urban hierarchy has been underexplored in analyses of pandemic-induced behavioral change. This study investigates how COVID-19 altered recreational gathering behavior in Seoul by leveraging motel booking data as a proxy for recreational activities. It assesses how recreational hierarchy, geographical distance, and attachment to familiar locations shape visitation patterns and how these factors shifted before and after the COVID-19 outbreak.

Prior research documents hierarchical organization of urban activity and mobility, hotspot dynamics, and their links to city structure, economic processes, and epidemic spread. Studies highlight that strong hierarchies can amplify disease transmission from popular hubs and exacerbate inequalities, particularly within informal economies. Mobility research shows universal visitation laws, hierarchical flows, and polycentric structures, while pandemic-era work links mobility reductions to transmission changes and unequal outcomes. However, explicit consideration of recreational hierarchy in behavioral responses to pandemics has been limited, motivating this study’s focus on hierarchy, distance, and place attachment in recreational gatherings.

Data: (1) Accommodation reservation data from Goodchoice Company LTD (29% online market share in 2020) include anonymized customer-level histories and geographic locations for 1038 unique motels in Seoul, spanning Jan 2019–Nov 2020; no demographics available. (2) Seoul mobility data (Jan–Oct 2020) from Seoul Open Data Plaza capture flows across 425 administrative divisions, disaggregated by gender, age, departure/arrival times, and classify locations as daytime residence (W), nighttime residence (H), and other (E). Non-routine mobility types used: WE, HE, EE to track recreational-related movements. Spatial discretization and hierarchy: Motels are assigned to Google S2 level-14 cells (~0.19–0.40 km^2). Reservation counts aggregated per cell define a recreational hierarchy via inverse decile ranks: level 1 (highest) to level 10 (lowest). Measures: (1) Hotspot entropy h = −∑ p_i log p_i over hotspot levels (L=10), where p_i is the frequency of level i in a customer’s level trajectory. (2) Radius of recreational activities r = sqrt(∑ d_i,home^2 / N), where d_i,home is the Haversine distance from the trajectory’s most frequent (home) cell and N is trajectory length. Study design: Periods defined as pre-COVID-19 (Jan 21–Nov 3, 2019) and post-COVID-19 (Jan 20–Nov 1, 2020), each 286 days. A focus group of customers with at least two reservations in both periods (30% of customers pre, 26% post) is analyzed for behavioral change. Individual records form cell and level trajectories; a flow matrix T_data counts normalized transitions between different hotspot levels (excluding self-transitions). Comparisons with aggregated mobility inflows validate bookings as a proxy. Mechanistic model: An agent-based generative process produces visit sequences matching empirical sequence lengths. At iteration i (starting at 1), the agent explores with probability p^i or returns to a previously visited place with probability 1−p^i via preferential attachment to past visit frequencies. If exploring, the agent samples a hotspot level from f(l) ∝ l^{−k} (k controls preference for higher hierarchy levels), then selects a place within that level with probability proportional to distance^{−γ} from the current recreational home (the most frequent cell so far); the home cell can update as history evolves. Parameter estimation uses grid search to minimize Jensen–Shannon divergence (JSD) between synthetic and empirical distributions: jointly optimize p and k to minimize JSD of hotspot entropy and hierarchy distributions (JSD_Entropy × JSD_Hierarchy), then optimize γ to minimize JSD of the radius distribution (JSD_Radius). Grids: p ∈ [0,1] in 51 bins (0.02), k ∈ [0,3] in 121 bins (0.025), γ ∈ [0,5] in 201 bins (0.025); 10 simulation repetitions averaged. Variant ablation models assess necessity of each factor: no hierarchy (k=0), no geography (γ=0), no attachment (i=1).

• Validation of proxy: Reservation counts per administrative division correlate with mobility inflows (Spearman ρ=0.347, p<0.001), stronger at night (ρ=0.400, p<0.001).
• Temporal impact: Weekly reservations drop sharply around the first week of official social distancing, then gradually recover despite restrictions.
• Recreational hierarchy stability and inequality: High-level cells (e.g., Gangnam, Sinchon, Yeongdeungpo Time Square) remain top hotspots across periods. A large share of reservations concentrate in the top 10% of cells (61.6% pre; 60.7% post), whereas the bottom 10% receive very few (0.6% pre; 0.3% post). After COVID-19, the share at the very highest level decreases, with relative increases at levels 2–4 and decreases at lower levels, indicating worsened inequality across urban areas.
• Individual activity: Among the focus group, average individual reservation counts decreased (pre mean 9.200 vs post mean 8.757; paired t=8.820, p<0.001). People explored fewer hierarchy levels (lower hotspot entropy) and maintained similar attachment to a recreational home (home ratio stable across sequence lengths).
• Flow patterns: Transitions concentrate among high hierarchy levels and are nearly symmetric. Model reproduces flow matrix with low Frobenius norm distance (~0.03).
• Model fit and ablations: Best-fit parameters pre vs post indicate reduced exploration and preference for top hotspots and increased distance aversion: p decreased from 0.820 to 0.800; k decreased from 2.075 to 2.025; γ increased from ~1.325 to ~1.375. Using pre-COVID parameters on post-COVID data worsens fits (JSD_Entropy×JSD_Hierarchy +7%; JSD_Radius +16%). Sensitivity is high: small deviations in p (±0.02), k (±0.025), or γ (±0.025) increase JSD metrics notably. Ablations show hierarchy is critical (k=0 collapses flows and entropy), geography determines spatial spread (γ=0 collapses radius), and attachment explains persistent home ratios (i=1 fails to reproduce retention).
• Behavioral interpretation and inequality: During the pandemic, individuals favored familiar, less popular, and closer places, likely reflecting avoidance of dense areas and reduced public transport use (bus −26.5%, subway −27.5% in 2020 vs 2019). Low-level urban areas experienced larger visitation declines than high and mid-level areas.

Findings show that COVID-19 shifted recreational behavior toward familiarity, lower hierarchy levels (but not the lowest), and proximity to recreational homes. This reduced concentration at the very top hotspots but simultaneously further suppressed activity at low hierarchy levels, worsening inequality across urban areas. The mechanistic model explains these dynamics through concurrent reductions in exploration propensity (lower p), decreased attractiveness of top-tier hotspots (lower k), and stronger distance deterrence (higher γ). These shifts are consistent with risk avoidance, social distancing, capacity limits, and reduced reliance on public transport. The results highlight how hierarchy, geography, and attachment jointly structure recreational gatherings and how external shocks reconfigure their balance, with implications for disease spread and urban economic resilience.

The study introduces a hierarchy-aware, distance-sensitive, and attachment-driven framework to quantify recreational visitation patterns using motel booking data in Seoul. It demonstrates that the COVID-19 pandemic led people to prefer familiar, less crowded, and nearby locations, altering flows across the recreational hierarchy and exacerbating spatial inequality by disproportionately reducing visits in low-level areas. The proposed model reproduces observed patterns and can serve as a simulation tool for policy planning under shocks. Future research should incorporate individuals’ pre-existing histories, networked transportation frictions, richer behavioral layers, and causal decomposition of voluntary versus policy-driven changes using higher-resolution data and advanced causal models to better inform equitable and effective public health and urban policy.

• Modeling initialization assumes agents start with empty reservation histories; real users possess prior habits that may shape choices.
• Geographic distance is measured as straight-line (Haversine) distance between S2 cell centers; transportation networks and travel times may distort effective distance.
• Motel bookings are a proxy correlated with mobility inflows; other recreational activities and contextual layers are not directly observed.
• Observed behavioral changes reflect both voluntary risk avoidance and policy constraints (social distancing, operating-hour and occupancy limits), which are not separately identified; high-resolution data and causal designs are needed for decomposition.
• Platform users are primarily low- and middle-income and skew younger and male, potentially limiting generalizability to the broader population.