Temporal dynamics have long been recognized as a fundamental dimension of human mobility, reflecting how daily activities, social obligations, and transportation systems shape when people move. Early mobility research emphasized the role of time in structuring travel behavior, demonstrating that human movement follows regular daily and weekly rhythms driven by work schedules, school hours, and service availability [10][14]. These temporal regularities have been widely used to identify peak travel periods, assess system performance, and inform transportation planning and demand management strategies.

Subsequent studies have leveraged increasingly detailed mobility data to examine temporal variability in travel behavior, including differences between weekdays and weekends, seasonal fluctuations, and deviations from routine patterns. Research using travel diaries and time-use surveys has shown that temporal constraints strongly influence trip timing, duration, and frequency, often limiting individuals’ ability to adjust their schedules in response to congestion or service disruptions [4][15]. More recent work utilizing large scale mobility datasets, such as mobile phone records and location-based services, has further revealed stable temporal signatures of human activity across cities, highlighting the predictability of collective movement patterns over time [5][11].

Temporal mobility patterns have also been linked to broader questions of accessibility and equity. Studies have shown that irregular or highly constrained temporal movement may indicate limited transportation options, inflexible work conditions, or restricted access to essential services, particularly for disadvantaged populations [16][17]. Peak-period congestion and off-peak service gaps can create temporal barriers that disproportionately affect individuals with limited schedule flexibility, reinforcing existing inequalities in access to opportunities.

Despite these advances, much of the temporal mobility literature focuses primarily on when movement occurs, with limited consideration of the spatial extent or distance associated with observed temporal patterns. Temporal analyses are often conducted independently of trip length or spatial reach, making it difficult to assess whether observed activity rhythms correspond to short, localized movements or longer-distance travel driven by service scarcity or spatial mismatch. As a result, while temporal dynamics provide valuable insight into the timing and regularity of mobility behavior, they offer an incomplete understanding of everyday movement when examined in isolation from distance-based considerations.

Distance-based mobility analysis has been widely used to examine the spatial extent of human movement and to assess how access to opportunities is shaped by the built environment and transportation systems. Early studies in transportation and urban economics emphasized trip length distributions and distance decay effects as fundamental characteristics of travel behavior, demonstrating that the likelihood of interaction between locations generally decreases with increasing distance [12][18]. These approaches have been central to understanding spatial interaction, land-use efficiency, and the organization of urban activity systems.

Subsequent research has expanded distance-based analysis to explore accessibility, spatial mismatch, and service availability. Trip distance has frequently been used as a proxy for spatial opportunity, with shorter trips often interpreted as indicators of compact urban form or constrained mobility, and longer trips associated with service scarcity, employment dispersion, or inefficiencies in land-use planning [9][19]. Accessibility-based studies have further linked travel distance to socioeconomic outcomes, highlighting how longer travel requirements can impose disproportionate burdens on certain populations [16][20].

With the emergence of large-scale mobility datasets, recent studies have examined distance-based movement patterns at unprecedented spatial resolution. Analyses using mobile phone data, GPS traces, and location-based services have revealed consistent distributions of travel distances across cities, as well as scaling relationships between trip length, urban size, and activity type [5][21]. These findings suggest that distance-based mobility behavior exhibits stable statistical properties at the population level, making it a valuable lens for comparative urban analysis and system-level evaluation.

Despite these contributions, distance-based mobility studies frequently abstract movement from its temporal context. Trip length and spatial reach are often analyzed independently of the timing, frequency, or regularity of travel, limiting the ability to distinguish between short-distance routine activities and longer-distance trips constrained to specific time windows. As a result, distance-based analyses alone provide limited insight into how spatial accessibility interacts with temporal constraints to shape everyday mobility behavior. This separation restricts a comprehensive understanding of population-level movement patterns, underscoring the need for integrated temporal-distance approaches.

Population-level mobility analysis has gained increasing attention as researchers seek to understand collective movement behavior while avoiding the limitations and privacy concerns associated with individual-level tracking. Rather than focusing on personal travel decisions, population-level approaches examine aggregated mobility patterns to reveal recurring rhythms, dominant travel behaviors, and systemwide dynamics that emerge from the interaction of individuals with transportation networks and urban form [6]. This perspective aligns closely with transportation planning and policy analysis, which prioritize demand patterns, peak loads, and accessibility conditions at the system level.

Early population-level studies relied on aggregated traffic counts and census-based travel indicators to infer spatial and temporal patterns of movement. While informative, these data sources were often limited in their ability to capture non-motorized travel, short-distance trips, and fine-grained temporal variability [4]. More recent advances in data availability have enabled population-level mobility analysis using large-scale digital traces, including mobile phone records, location-based services, and foot traffic datasets. These sources provide continuous observations across extended periods, allowing researchers to identify stable mobility signatures, compare activity patterns across locations, and assess changes in movement behavior over time [5][7].

Aggregated mobility data have been widely applied to study urban activity patterns, land-use interactions, and accessibility dynamics. Studies have shown that population level movement exhibits regular statistical properties, such as predictable temporal cycles and consistent distance distributions, despite individual-level variability [11][21]. By focusing on collective behavior, population-level analysis can uncover structural constraints such as limited-service availability or transportation capacity that are not easily detectable through individual travel records alone [9][16].

Network analysis is a potent tool to explain mobility, as it treats locations as nodes of a network and movements between them as weighted edges that can be quantified within the framework of complex systems [22]. The connectivity of mobility networks, the extent to which areas are connected through the movement of humanity, shines brightly as an important determining factor of local and regional dynamics [23]. Centrality measures identify the position of a location within networks and highlight those hubs that are at the center of human flows [24], where people tend to aggregate and disperse in a highly evident, unequal manner. Temporal networks, i.e., networks based on human mobility, show that the topology of movement between places is highly heterogeneous across time of day, seasons, and other external perturbations, undermining fixed, static network models [25]. The two-dimensional spatio-temporal approach that brings together temporal and distance-based dimensions into network frameworks shows how physical accessibility and temporal regularities together govern human movement patterns [26].

Spatial coupling, that is, the extent to which spatial proximity and structural network properties interact to produce mobility outcomes, represents an emergent outcome, one that cannot be understood from either spatial terms or the characteristics of the interrelations represented in networks alone [24]. Latest studies of mobility networks in various contexts show that connectivity and centrality are not isolated properties but rather closely linked with a spatial dimension of urban areas, the accessibility to infrastructure or land use organization, as well as socioeconomic tendencies at places [27]. Human mobility networks are inherently multilayer and temporal in nature, making it necessary for centrality measures to consider the structural roles of locations within static network topologies as well as their dynamic importance across different temporal scales [28]. The spatial coupling dynamics arise most clearly in the context of storms and other disruptive events: mobility networks reorganize structurally over extremely short time scales, as long-distance connections are increasingly broken while local clustering increases to a greater or lesser extent, with fundamental implications for how we characterize connectivity and centrality metrics [29]. Hence, to sensitively model large-scale phenomena like epidemic spreading through meta-population networks, it is essential to recognize these three interwoven facets of mobility: connectivity, centrality, and spatial coupling [30].

Despite these methodological advances, relatively few studies have leveraged aggregated mobility data to jointly examine temporal and distance-based dimensions of movement within a unified analytical framework. Existing population-level analyses often emphasize either temporal regularity or spatial extent independently, rather than exploring how the interaction of when and how far people travel reflects accessibility conditions and mobility constraints. As a result, the potential of population-level data to provide an integrated understanding of everyday mobility behavior remains underutilized. Addressing this gap requires analytical approaches that simultaneously capture temporal dynamics and distance-based patterns to reveal the structural characteristics of population-level movement systems.

The study utilizes the Weekly Patterns+ foot traffic dataset from Dewey, a SafeGraph-derived mobility dataset that offers aggregated and anonymized insights into human mobility patterns. It tracks activities from Monday to the end of Sunday in each period. Available from January 1st, 2018, to June 30th, 2025, the dataset covers 50 states across the United States and reports mobility at the Census Block Group (CBG) and point of interest (POI) levels. These 50 states include: Alaska, Arizona, Arkansas, Alabama, Washington, Oregon, Montana, Idaho, Nevada, California, New Mexico, Texas, Florida, Mississippi, Oklahoma, Utah, Colorado, Wyoming, Nebraska, South Dakota, North Dakota, Kansas, Louisiana, Georgia, Massachusetts, North Carolina, Delaware, Maryland, Rhoda Island, Virginia, Ohio, Iowa, Minnesota, Indiana, Tennessee, Kentucky, West Virginia, New York, New Jersey, South Carolina, Maine, New Hampshire, Wisconsin, Michigan, Illinois, Vermont, Pennsylvania, Connecticut, Hawaii, Missouri.

The data was derived from mobile device location pings collected through opt-in applications. Each observation level represents the aggregated weekly total visits and stops at the point of interest (POI) levels to preserve user privacy. In addition, the analytical sample was constructed by including POI-week observations with full weekly visitation information, valid temporal activity distributions, and non-missing distance-based mobility measures. Records were excluded from the analysis if they had incomplete geographical identifiers, an incomplete distribution for hourly visitation, duplicate observations, or no activity at all in the location. This structure enables the analysis of temporal and distance-based human mobility patterns across diverse geographic states in the United States.

2018: ≈2 million records2019: ≈2 million records 2020: ≈2 million records 2021: ≈2 million records 2022: ≈2 million records 2023: ≈2 million records 2024: ≈2 million records 2025: ≈1 million records

Each dataset includes essential foot traffic attributes such as visit counts, stop frequency, temporal activity distributions by hour and day, stops by day, dwell-time indicators, device type information, and spatial identifiers.

The analytical framework adopted in this study follows a structured four-stage workflow illustrated in Figure 1. Beginning with data collection and preprocessing, the framework progresses through feature engineering, modelling and analysis, and evaluation and interpretation. The modelling and analysis stage encompasses four complementary components: spatiotemporal analysis, network analysis, distance modelling via OLS regression, and behavioural modelling via the Multinomial Logit model. Each stage builds sequentially on the outputs of the preceding stage, ensuring analytical consistency across temporal, spatial, and behavioral dimensions of population-level mobility behavior.

Figure 1. Proposed integrated workflow.

Before model development, comprehensive preprocessing was undertaken on the dataset to ensure that the data is consistent, accurate, and ready for analysis. Temporal data is standardized to make the date columns consistent, create new features for year, month, week, and day of the week, etc., for the dataset.

The raw dataset was originally stored in individual files for each month of each calendar year, so first of all, it is necessary to merge all of them to create complete annual data sets, after which they needed to be concatenated in order to create one united data set of mobility data ranging from 2018 to 2025.

Nested and semi-structured data for temporal activity patterns, recorded as JSON format strings for daily foot traffic counts, are decoded to structured numeric format, for example, “Monday”: 5, “Tuesday”: 5, “Wednesday”: 5, … Daily foot traffic counts are then segregated to individual weekday counts, which allow for the analysis of foot traffic patterns for specific weekdays, i.e., weekends, as well as other specific periods of the week. Also, the individual brand-level foot traffic counts recorded in JSON format, for example, “Walmart”: 24, “McDonald’s”: 9, “Hospitals”: 5, are decoded to individual brand-level counts, with missing records being treated as empty records, to allow for the analysis of foot traffic patterns for individual brands, including the ones for which the foot traffic counts are missing or invalid.

Feature engineering was a vital step in developing the analytical framework to analyze temporal as well as distance-based patterns of human mobility using aggregated foot traffic datasets. The aim was to identify and construct features that are effectively captured when individuals travel, how far they move from their homes, and how their activity patterns change with location and time, while ensuring consistency across regions and years.

1) Brand to Category Mapping: The mapping between brand and category was performed using a hybrid rule based classification framework across individual brands and points of interest (POIs), which were segmented into higher-level functional activity categories such as: Retail, Restaurant, Healthcare, Automotive Dealerships, Recreation, Gas Station, Lodging, Transportation & Logistics, Education, Government & Public Services, Financial Services, Personal Care Services. This included lowercasing text, removing punctuation, whitespace normalization, and harmonization of keywords to standardize brand names. Classification rules were mostly based on keyword matching along the lines of industry semantics and business functionalities. Brands related to supermarkets, department stores, and convenience stores would be assigned Retail; hospitals, clinics, and pharmacies Healthcare; restaurants, cafés, and fast-food stores Restaurant; freight companies, airports, and places where people assembled for travel operations and delivery services to Transportation & Logistics.

In addition, to derive the mobility intensity profiles for each activity category, hourly stop count aggregates h00 to h23 are used to identify peak and non-peak movement times and to account for temporal variation in activity participation.

Ambiguous POIs, which have reasonable claims to belong to more than one category, were resolved through a hierarchical decision process. I began by identifying the predominant operational function of the POI using brand description and contextual business activity. If any uncertainty remained, the POI was assigned to represent its main consumer-facing function; for instance, a retail store containing a pharmacy section was categorized as Retail unless healthcare services dominated. POIs with insufficient description, conflicting codes, or overly generic names were assigned to an “Other” category to avoid any classification bias or the inflation of numbers in certain categories.

2) Feature Screening and Selection: Each dataset had a multitude of features that described mobility patterns as well as other contextual attributes, such as weekly visit counts, stops made by individuals daily as well as by hour, day of week activity distribution, distance from home, dwell times, device origin, as well as brand-level visit data, among others. To understand the interdependence between these features and their relationship with mobility outcomes, an exploratory correlation analysis was performed using Pearson correlation coefficients. Highly collinear variables with absolute correlation coefficients exceeding |0.90| were considered redundant, and one representative variable was retained based on interpretability and completeness. Based on this correlation, strong correlations were observed between overall visit intensity, as well as temporal features such as weekday-weekend patterns, peak day activity, and visit behavior, as well as distance from home features, indicating the importance of considering these two dimensions of mobility. Variables with near-zero variance, excessive missingness, or constant values across observations were removed as they contributed minimal explanatory information. Categorical features with extremely sparse representation were consolidated into broader groups or excluded where appropriate to avoid unstable estimates during modeling.

Feature relevance was further evaluated through the Multinomial Logit modeling framework, which assessed the stability and interpretability of retained features across all 18 activity categories. Results consistently confirmed that temporal features demonstrated the strongest and most consistent contributions to differentiating mobility behavior across destination types. Distance-based mobility behavior was examined separately through an OLS regression model in which distance from home served as the continuous dependent variable, with activity category, U.S. state, and year as categorical predictors. Full model results for both the Multinomial Logit (MNL) and Ordinary Least Squares (OLS) specifications are reported in the Results section.

Using the results of the feature evaluation, it has been possible to develop a new set of features that are most relevant and suitable to understand and analyze human mobility behaviors. The new set of features that is most relevant and suitable to understand and analyze human mobility behaviors is as follows:

Temporal features that help in understanding and analyzing the rhythms of movement and distinguish between weekday and weekend movements.Distance-from-home features that help in understanding and analyzing the constraints of accessibility and reach of daily activities.Brand-level activity features derived from the category mapping framework to differentiate mobility behavior across functional destination types.

3) Aggregation and Sample Reduction: Following the brand-to-category mapping and feature construction procedures outlined above, individual brand-level records were grouped into 18 functional activity categories, which served as the unit of analysis for behavioral modeling. The raw dataset was at the individual brand level and contained around 15 million POI-week records, each observation indicating mobility activity in a specific named brand or point of interest on a weekly basis. This aggregation collapsed the brand-specific units of observations into weekly mobility profiles at the category-level, resulting in an analytical sample of 180,913 category-level observations down from more than 15 million brand-level records. Aggregated across the analytical period: 2018-2025, each observation in the data identifies total weekly mobility activity for a given functional activity category. This category-level structure forms the unit of observation for the Multinomial Logit model mentioned, ensuring that the behavioral model operates on stable, consistently defined mobility profiles rather than noisy individual brand-level records.

The results were estimated using maximum likelihood estimation (MLE) in the stats models library. Model convergence was established after 8 iterations, and statistical significance was evaluated at the α = 0.05 level throughout. An OLS regression model was estimated to examine the spatial determinants of distance-based mobility behavior, with distance from home serving as the continuous dependent variable. Each observation represents a POI-week record capturing the weekly aggregated mobility activity associated with a specific point of interest. Activity category, U.S. states, and year were entered as categorical predictors using dummy coding, with Automotive Dealerships as the reference activity category, Alaska as the reference state, and 2018 as the reference year. The OLS model was fitted on 256,519 POI-week observations using heteroskedasticity-robust standard errors (HC3 covariance type) to account for non-constant variance across the large observational sample. The likelihood ratio test was used to evaluate the joint significance of the MNL model, and F-tests were used for the OLS model.

The category-level Anchoring Index was calculated using visit-weighted distance measures, differentiating between residential and daytime anchor contributions during weekday and weekend periods. Temporal decomposition was used to identify peak and off-peak mobility behaviors across hours of the day and days of the week. All modeling and index construction procedures were consistently applied across all 50 states and all years from 2018 through 2025 to ensure regional and temporal comparability. For the MNL, pseudo R² was reported alongside likelihood ratio testing for joint significance of predictors. For the OLS, R² and adjusted R² were used to assess model fit. Robustness of findings was evaluated through cross-regional consistency checks to verify the geographical stability of results across states.

Statistical computations, performance evaluation, and visualization were performed using numpy, pandas, matplotlib, and dask packages. The latter was selected for its ability to handle out-of-memory computation given the large analytical sample spanning approximately 15 million raw records across 2018-2025.

This study proposes an anchoring-based mobility modeling framework to characterize variations in human population-level temporal/distance-based movement patterns. The framework combines a formally defined Anchoring Index, visit-weighted distance aggregation, temporal decomposition (weekday vs. weekend and peak vs. off-peak), and network-based representations to measure the extent to which activity categories are spatially anchored in residential and daytime locations. Instead of assuming that sample choice models can predict individual-level mobility behavior, the framework uses aggregated mobility data to extract structural patterns of mobility across categories, regions, and time, yielding interpretable metrics such as spatial dependence, the burden of dislocation, and network connectivity in large-scale human movement systems.

1) Anchoring Index Formulation: To measure how population mobility organizes space around residential and daytime activity centers, we propose an Anchoring Index (AI) that characterizes the relative strength of home-based versus daytime-based travel for various kinds of activity.

Let each mobility observation aggregate visitation from a population of devices to destination category c, with distances associated with two reference anchors:

The residential anchor (home location), andThe primary daytime anchor (typically associated with work or routine daytime activity).

For each activity category c, we compute visit-weighted average distances:

where d_home(i) denotes the distance traveled from the home anchor, d_day(i) denotes the distance traveled from the primary daytime anchor, w_i represents visit intensity (weighted stop counts), and the summation is taken over all observations i associated with activity category c.

The Anchoring Index for category c is then defined as:

This formulation yields a continuous anchoring metric with an intuitive interpretation where,

AI_c > 0 indicates activity categories primarily anchored around residential locations,AI_c < 0 indicates activities more strongly tied to daytime or work-related anchors.AI_c ≈ 0 indicates spatially balanced or flexible activity patterns.

2) Temporal Decomposition of Anchoring: To capture temporal variation in spatial anchoring, the Anchoring Index was computed separately across multiple temporal regimes: Weekday vs Weekend.

Weekday visits reflect routine commuting and work-related mobility.Weekend visits capture discretionary and leisure-driven travel.

Peak vs Off-Peak Hours

Peak hours correspond to typical commute and activity surges.Off-peak hours reflect flexible, non-routine travel behavior.

For each temporal subset, category-specific Anchoring Indices were recalculated using the same weighted formulation, allowing direct comparison of how anchoring strength shifts across time.

3) Category-Level Distance Profiling: Using the Anchoring Index as the foundational metric, each activity category was further characterized by:

Weighted average distance from homeWeighted average distance from daytime anchorTotal visitation volume

This produces a distance profile for each category, enabling differentiation between:

Local routine activities (short distances, strong home anchoring),Work-oriented activities (negative anchoring, daytime centric),Discretionary activities (long distances, weak anchoring).

4) Network Construction and Analysis: Based on the aggregated foot traffic dataset, a bipartite category-region mobility network was constructed to explore structural relationships between activity categories and geographic regions. Nodes represent functional activity categories, or U.S. states, and each state is connected to every category when visits were recorded in that state for multiple participants. Edge weights were defined as the sum of visit volume aggregated across a category/state pair over the entire 2018-2025 analytical period. This resulted in a 75-node, 1043-edge network (density = 0.376) with moderate to high connectivity throughout the category-region network. I computed three network measures for each category node: a) degree centrality, quantifying the proportion of states connected to each category; b) betweenness centrality, which measures the extent to which a category node lies on shortest paths between other nodes, reflecting its bridging role in the network; and c) node strength, or total weighted visit volume across all connected states. Spatial coupling was assessed through the coefficient of variation (CV) of distance-from-home measures across states for each category, where elevated CV values reflect greater spatial dispersion in visitor origin, while lower values reflect spatially concentrated or locally anchored visitation patterns.

5) Behavioral Modeling (Multinomial Logit): Category-level mobility behavior was formally modeled through a Multinomial Logit (MNL) model with two groups of predictors. The unit of observation is a POI-week record, which are records of the aggregated mobility activity at each point of interest in a particular week. As for the outcome variable, it is a nominal functional activity family of the destination visited, which has 18 mutually exclusive mobility categories—Automotive Dealerships, Commercial & Trade Services, Community & Religious Services, Education, Financial Services, Funeral Services, Gas Station, Government & Public Services, Healthcare, Lodging, Mall, Other Personal Care Services, Pet Services, Recreation, Restaurant, Retail, Transportation & Logistics.

Three factors explain why the MNL framework is suitable for this analysis. This is first the dependent variable is nominal, while there is no natural ordering across activity categories that would meet the foundational requirement of MNL over ordered alternatives. Second, the aggregated population-level foot traffic records on which the model operates, rather than individual-level discrete choices, are especially useful for characterizing systematic behavioral patterns across large observational samples. Third, by using a class of conditional logit models with category-specific coefficient structure (as in the MNL), we can directly compare the effects of temporal sensitivity and activity intensity for each of the 18 mobility categories together and preserve interpretability at the population level.

The model included three fixed explanatory variables: hour-of-day, capturing the linear time-of-day effect on category-level mobility; hour², a quadratic term to capture non-linear temporal patterns that show their peak over midday and then dropping in the evenings; log-transformed total stops, normalized on a log-scale to reflect overall activity intensity. Maximum likelihood estimation: Converged after 8 iterations. (183,913 observations on 18 features); of class “hlm” Joint statistical significance is confirmed by the likelihood ratio test (LLR p < 0.001), and a pseudo-R² of 0.093 is observed, in line with expectation for population-level aggregated mobility data whilst controlling for individual-level heterogeneity. This amount of explanatory power is permissible for MNLs examining multi-category behavioral outcomes in large/heterogeneous observational samples.

Table 1 presents coefficient estimates, standard errors, z-statistics and 95% confidence intervals for all 18 relatively non-reference categories; coefficients are estimated with high precision across all categories (standard errors range from 0.006 to 0.113), reaffirming the statistical stability possible with such a large analytical sample. In most categories, all three predictors are statistically significant at the α = 0.05 level. Two categories show no significant temporal coefficients (Community & Religious Services: hour: p = 0.077, hour²: p = 0.123; and Healthcare, hour: p = 0.517, hour²: p = 0.570). And these results are interpretable—Community & Religious Services visits cluster on days of the week (weekends vs. weekdays) rather than hours of a day, and Healthcare visits cluster around scheduled appointments and medical needs, which are negligible determinants for hour-of-day choice, letting time-of-day be a relatively weak predictor for both destination types. Table 1 reports 95% confidence intervals alongside standard errors and z-statistics for all category specific estimates showing that the effects reported in our main plots are stable and precise, as reflected in small 95% confidence intervals.

Table 1. Multinomial logit model estimates for category-level mobility behavior.

SE = standard error. ^†Indicates statistically insignificant categories at α = 0.05. N = 180,913 POI-week observations. Pseudo-R² = 0.093. LLR p-value < 0.001.

Population-level mobility derived from the 2018-June 2025 foot traffic dataset exhibits strong temporal regularity and meaningful regional variation across activity categories. The average hourly mobility profile, shown in Figure 2 (Average Hourly Mobility Profile by Category), reveals a pronounced diurnal structure. Across nearly all categories, activity levels remain low during early morning hours (00:00-05:00), increase steadily beginning around 07:00, and peak between approximately 15:00 and 17:00. Following the late-afternoon peak, mobility gradually declines into the evening. This inverted U-shaped pattern indicates that human movement is highly structured within predictable daily rhythms rather than randomly distributed throughout the day.

Figure 2. Average hourly mobility profile by category.

Despite the shared diurnal structure, substantial heterogeneity exists across categories. Retail, Restaurant, and Mall exhibit the highest overall intensity and the most pronounced afternoon peaks, suggesting strong alignment with commercial and leisure-oriented activity cycles. In contrast, Education and Healthcare demonstrate narrower activity windows and relatively earlier stabilization during the afternoon. Categories such as Community & Religious Services and Funeral Services show comparatively lower hourly variation, indicating weaker sensitivity to hour-of-day dynamics. These differences confirm that temporal mobility behavior varies systematically according to activity type.

Long-term temporal trends further highlight category-level divergence. As illustrated in Figure 3 (Category Mobility Trends Over Years, 2018-2025), several high-intensity categories particularly Retail, Mall, and Restaurant display marked growth between 2022 and 2024, reflecting substantial post-2022 expansion in commercial mobility activity. Health-care and Commercial & Trade Services also demonstrate upward trajectories during this period. The observed decline in 2025. across several categories likely reflects partial-year data coverage rather than structural contraction. Overall, these multi-year patterns indicate both recovery and expansion dynamics in key economic sectors.

Figure 3. Category mobility trends over years (Top 10 categories).

In addition to temporal variation, mobility intensity differs significantly across regions. Figure 4 (Mobility Intensity by Selected Categories OverTime: CA, NC, TX, KS, CO, FL) compares total stop volumes across six states. Texas and Florida consistently exhibit higher aggregate mobility intensity relative to North Carolina and Kansas, while California shows moderate but stable growth over time. Although the general temporal trajectory remains broadly similar across states, absolute activity levels vary considerably, indicating that regional context plays a significant role in shaping overall mobility intensity.

Figure 4. Mobility intensity by selected categories over time (CA, NC, TX, KS, CO, FL).

Taken together, Figures 1-3 demonstrate that human mobility exhibits strong daily temporal regularity, sustained multi-year growth in key commercial categories, and meaningful regional heterogeneity. These descriptive spatiotemporal patterns provide the empirical foundation for subsequent analyses examining activity intensity, spatial distance characteristics, and behavioral determinants of category-level mobility outcomes.

Beyond hourly and annual dynamics, overall activity intensity and weekly structure provide additional insight into category-level mobility behavior. Figure 5 presents a comparison of average weekly visits by category, disaggregated into weekday and weekend totals. Clear differences emerge across activity types, indicating that mobility is not uniformly distributed across the week.

Education exhibits the highest overall weekly volume, with a strong dominance of weekday activity, reflecting institutional schedules and structured daily routines. Mall and Retail also show substantial total visits, but with a comparatively larger weekend contribution, consistent with discretionary and leisure-oriented shopping behavior. Restaurants demonstrate high activity across both weekday and weekend periods, suggesting its dual role in routine daily consumption and social or recreational activity.

Figure 5. Weekdays vs weekend activity by category.

Categories such as Gas Station and Financial Services display relatively stable weekday-dominated structures, reflecting essential and transaction-based mobility needs. In contrast, Recreation and Travel exhibit comparatively higher weekend shares, indicating discretionary temporal flexibility and non-work-related mobility.

The relative magnitude of category-level activity is further illustrated in Figure 6, which visualizes mobility intensity weighted by total daily stops. Larger word sizes correspond to higher aggregate mobility contributions over the study period.

Figure 6. Mobility intensity by category (Weighted by total daily stops).

As shown in Figure 6, Retail, Mall, and Restaurant dominate overall mobility intensity, reinforcing their central role in shaping aggregate foot traffic patterns. Automotive Dealerships and Healthcare also contribute substantially, while categories such as Funeral Services and Community & Religious Services represent comparatively smaller shares of total mobility. This concentration of activity within a limited set of high-intensity categories suggests that a small number of economic sectors account for a disproportionate share of population-level movement.

Together, the weekly distribution and intensity patterns indicate that mobility behavior reflects both structured institutional schedules and discretionary consumer activity. Categories with strong weekday dominance (e.g., Education and Financial Services) reflect routine-based movement, whereas categories with elevated weekend shares (e.g., Recreation and Travel) reflect flexible or leisure-driven mobility. These distinctions provide important context for understanding how temporal dynamics interact with spatial and behavioral determinants examined in subsequent sections.

While temporal dynamics describe when mobility occurs, spatial distance characteristics reveal how far individuals travel and whether activities are locally anchored or spatially dispersed. The distribution of travel distances from home across categories is presented in Figure 7 (Distance from Home by Category).

Figure 7. Distance from home by category.

As shown in Figure 7, most categories exhibit highly right skewed distance distributions, with the majority of visits occurring within relatively short distances from home but with occasional long-distance outliers. Categories such as Travel, Lodging, and Recreation display wider dispersion and higher extreme values, indicating greater spatial reach and longer-distance mobility behavior. In contrast, categories including Gas Station, Financial Services, and Personal Care Services demonstrate tighter clustering around shorter distances, suggesting localized, routine-oriented activity.

Retail, Restaurant, and Mall occupy an intermediate position. While a substantial proportion of visits occur close to home, these categories also exhibit moderate spatial dispersion, reflecting both neighborhood-level and destination oriented trips. This pattern suggests that commercial mobility combines localized convenience-based visits with longer distance, centralized activity hubs.

To further evaluate spatial anchoring behavior, Figure 8 (Home vs Daytime Anchoring by Category) compares the average difference between home-based and daytime locations across categories. Positive values indicate categories more closely associated with home-origin activity, whereas negative values reflect stronger association with daytime or work-based spatial contexts.

Figure 8. Home vs daytime anchoring by category.

The anchoring analysis reveals distinct spatial roles across categories. Community & Religious Services and Education exhibit strong positive home anchoring, indicating that visits are typically concentrated within residential proximity. Government & Public Services and Commercial & Trade Services also demonstrate moderate home association.

In contrast, Travel and Pet Services show pronounced negative anchoring values, suggesting stronger alignment with daytime or non-residential spatial contexts. Restaurant and Retail appear closer to neutral anchoring, reflecting their dual function as both neighborhood-based and destination based activities. Mall and Lodging similarly display balanced or mildly negative anchoring, consistent with centralized commercial or hospitality destinations.

Collectively, these spatial findings indicate that mobility categories differ not only in intensity and timing but also in spatial reach and anchoring structure. Routine-service categories tend to exhibit localized, home-proximate behavior, whereas discretionary or destination-based categories display broader spatial dispersion. These distance-based distinctions complement the temporal patterns identified earlier and provide critical context for interpreting the behavioral modeling results presented in the following section.

Ordinary Least Squares Model: To examine the spatial determinants of mobility behavior, an OLS regression model was estimated with distance from home as the continuous dependent variable; full model specification details are provided in Section III (Experimental Setup). The model achieves an R² of 0.086, indicating that category, region, and year jointly explain approximately 8.6% of the variance in distance-from-home behavior, with the F-statistic confirming strong joint statistical significance (p < 0.001). While the proportion of explained variance is modest, this is consistent with expectations for population-level aggregated mobility data, where substantial within-category heterogeneity remains unexplained by category membership alone. Full coefficient estimates are reported inAppendix.

In Appendix, results reveal substantial heterogeneity in distance-from home behavior across activity categories relative to the Automotive Dealerships reference category. Lodging exhibits the largest positive distance effect (β = 21,940, SE = 1740, p < 0.001), followed by Transportation & Logistics (β = 8325, SE = 2908, p = 0.004) and Gas Station (β = 4771, SE = 1232, p < 0.001), confirming that these destination types attract visitors from the greatest distances relative to the reference category. In contrast, Funeral Services (β = −9192, SE = 1958, p < 0.001) and Personal Care Services (β = −7033, SE = 1205, p < 0.001) exhibit the largest negative effects, reflecting their strongly proximity-driven, residentially anchored demand patterns. Several categories, including Mall, Restaurant, Retail, Financial Services, Education, Government & Public Services, and Community & Religious Services, show no statistically significant difference in distance-from-home behavior relative to Automotive Dealerships (p > 0.05), suggesting broadly similar spatial reach across these destination types.

Temporal trends reveal a consistent increase in distance from-home mobility across years relative to the 2018 baseline. The largest positive deviation is observed in 2024 (β = 15,790, SE = 1044, p < 0.001), followed by 2025 (β = 12,080, SE = 899, p < 0.001) and 2023 (β = 6361, SE = 745, p < 0.001), indicating a sustained post-pandemic expansion in mobility reach across activity categories. Notably, 2020 shows no statistically significant change from the 2018 baseline (β = −231, SE = 587, p = 0.695), consistent with the severe mobility restrictions associated with the COVID-19 pandemic that effectively suppressed distance-from-home travel to pre-study levels.

To formally quantify the determinants of category-level mobility behavior, a multinomial logit (MNL) model was estimated using hour-of-day, a quadratic time term (hour 2), and log-transformed total stops as explanatory variables. The model was fitted on 180,913 observations and converged successfully under maximum likelihood estimation. The likelihood ratio test indicates strong joint statistical significance (p < 0.001), and the model achieves a pseudo-R² of 0.093, reflecting meaningful explanatory power given the multicategory structure and large alternative set.

1) Temporal Structure: The temporal coefficients reveal systematic heterogeneity in time-of-day sensitivity across mobility categories. A large subset of categories including Commercial & Trade Services, Education, Financial Services, Government & Public Services, Mall, Personal Care Services, Retail, and Travel exhibit a positive linear hour effect combined with a negative quadratic term. This sign pattern indicates an inverted U-shaped relationship, consistent with pronounced afternoon peaks observed in the descriptive hourly profiles. These categories demonstrate strong alignment with structured daytime activity cycles and institutional schedules.

In contrast, categories such as Automotive Dealerships, Funeral Services, Lodging, Recreation, and Restaurant display a negative linear hour coefficient paired with a positive quadratic term. This configuration reflects a U-shaped or extended temporal pattern, suggesting relatively stronger activity during early or later portions of the day compared with midday concentration. Restaurant, in particular, exhibits one of the strongest nonlinear effects, indicating substantial deviation from conventional business hours concentration.

Healthcare and Community & Religious Services show comparatively weak or statistically insignificant temporal curvature, indicating more stable distributions throughout the day. These categories appear less governed by sharp diurnal peaks and more by routine or necessity-based demand.

2) Sensitivity to System-Wide Activity Intensity: The results indicate marked variation in intensity sensitivity across categories (Table 1). Restaurant exhibits the largest positive log-transformed total stops coefficient among all categories (β = 0.926, SE = 0.009, p < 0.001), followed by Lodging (β = 0.516, SE = 0.008, p < 0.001) and Recreation (β = 0.417, SE = 0.008, p < 0.001), demonstrating that these categories expand disproportionately during high-mobility periods. These activity types appear demand-amplified, gaining relative share when aggregate movement intensifies.

Conversely, Financial Services, Other, Retail, Mall, and Commercial & Trade Services display strong negative coefficients on log (total stops). This does not imply absolute decline during busy periods, but rather a relative redistribution effect: when overall mobility increases, intensity-driven categories (notably Restaurant and Recreation) capture a larger share of total activity. These business-oriented categories are thus more scheduled-driven than intensity-amplified.

Moderate positive effects are observed for Automotive ealerships, Funeral Services, and Healthcare, suggesting partial responsiveness to aggregate activity without the amplification observed for discretionary sectors.

3) Distinct Behavioral Patterns Across Mobility Categories: Taken together, the temporal and intensity effects reveal clear heterogeneity in category-level mobility behavior.

a) Schedule-Driven Categories: Education, Financial Services, Government & Public Services, Retail, and Mall exhibit strong afternoon peaks and comparatively weaker amplification under high system activity. These categories are primarily structured by institutional or business-hour routines.

b) Demand-Amplified Categories: Restaurant, Lodging, and Recreation show strong positive intensity sensitivity and nonlinear temporal patterns. These categories disproportionately expand during high-mobility periods and reflect discretionary or consumption-driven behavior.

c) Routine-Stable Categories: Healthcare and Community & Religious Services display limited curvature and modest intensity responsiveness, indicating more necessity-based and spatially anchored mobility.

d) Empirical Coherence Between Descriptive and Modeling Results: The modeling results align closely with the descriptive spatiotemporal patterns presented in Sections 4.1 - 4.3. Categories identified as intensity-amplified correspond to those exhibiting strong weekend shares and broader spatial dispersion, while schedule-driven categories mirror the structured afternoon peaks observed in the hourly mobility profiles. This consistency reinforces the robustness of the analytical framework and confirms that temporal dynamics and aggregate system intensity jointly shape category-level mobility behavior.

1) Network Structure and Category Centrality: The systematic structural differentiation across activity types in connectivity, centrality, and visit strength is exposed by the category-region mobility network. Degree centrality, betweenness centrality, and node strength are listed for all 18 activity categories in Table 2.

Table 2. Category-region network centrality measures.

Note: Degree centrality is normalized by the total number of nodes in the category-region mobility network. Betweenness centrality is weighted by aggregated visit volume. Node strength represents the total aggregated visit volume across all connected states from 2018 to 2025.

Retail has the highest betweenness centrality (0.941) across all categories, meaning it is the largest bridging node in the category-region network; it connects essentially every state with a high volume of visits through retail as an entry. This is mirrored in the structure of Retail, the most ubiquitous type of Activity around different geophysical contexts. Healthcare (betweenness = 0.027) and Mall (betweenness = 0.0041), on the other hand, identify secondary bridging categories; while their visit frequencies are relatively low overall, they remain structurally significant in connecting diverse regional contexts. Most of the other categories have marginal values for betweenness centrality in a way that endorses this expectation (most of their regional connectivity lies in coarse, non-bridging spatiotemporal demand).

For node strength, Retail exhibits the highest weighted visit volume at 4.82 × 10¹⁰, followed by Mall (2.10 × 10¹⁰), Restaurant (1.17 × 10¹⁰), Automotive Dealerships (1.11 × 10¹⁰), and Commercial & Trade Services (1.09 × 10¹⁰). These categories represent high-frequency and geographically dispersed mobility demands and collectively account for the majority of total network visit flow.

In contrast, Personal Care Services (1.30 × 10⁸), Community & Religious Services (3.84 × 10⁸), and Government & Public Services (5.09 × 10⁸) exhibit substantially lower network strength values, reflecting their more localized and spatially bounded service functions within the human mobility network.

2) Spatial Coupling Analysis: The spatial coupling results in Lodging identifies a notable difference in the extent of visitor origins associated with each activity category Table 3.

Table 3. Distance-based mobility characteristics by category.

Note: Mean Distance and Std Distance are reported in meters. Spatial Coupling CV represents the coefficient of variation describing the relative spatial variability of mobility patterns across categories.

Education has the highest spatial coupling CV (12.50) as the distance to home of visitors to this category varies widely between states, likely reflecting both local schools and long distance university destinations that share a common name space within the same category label. Following this, Retail (CV = 10.14) and Mall (CV = 9.42) form a similar pattern, as both types appear to serve the same functional need at the neighborhood level but also play a role as a regional destination type due to their geographic contextualization.

Both Lodging(mean = 39,719 m) and Community & Religious Services(mean = 24,799 m) have high mean distances likely related to their connection with overnight and long distance mobility. On the other hand, Personal Care Services (mean = 8985 m), Gas Station(mean = 9328 m), and Funeral Services (mean = 7494 m) have the lowest mean travel distances from home rank swell in line with their strongly residential-anchor dependent proximity-driven demand patterns.

The combined findings on category-region mobility flows reveal that the structural differentiation of connectivity and spatial reach is not limited to network centrality measures, but extends to both dimensions of category-region mobility flows; in particular, high-frequency and essential categories dominate network strength whereas discretionary orregional specialty categories exhibit greater variability in visitor origins at a given market share.