Do Railway Commuters Exhibit Consistent Route Choice Rationality Across Different Contexts and Time? Evidence from Tokyo metropolitan Commutes

TL;DR

This study analyzes one year of high-resolution GPS trajectories from over a million Tokyo-area commuters to assess how consistently route choices are rational across contexts. By framing route choice within a canonical ensemble, it estimates a global energy cost combining time, distance, and transfer attributes, with the inverse temperature $\beta$ quantifying rationality. Findings show aggregate route-choice behavior is temporally stable and predominantly driven by non-monetary costs, while OD-specific heterogeneity reveals bimodal patterns and peak/off-peak differences, driven largely by transfer dynamics. The work demonstrates a scalable, data-driven transfer-aware approach to urban route-choice modeling, offering policy-relevant metrics like a transfer-penalty around $7$ minutes and guidance for capacity-focused transfer design. Overall, the canonical ensemble framework provides a principled lens to study collective mobility and informs more accurate, context-aware transport planning in dense networks.

Abstract

Recent advances in data collection and technology enable a deeper understanding of complex urban commuting, yet few studies have rigorously analyzed the temporal stability and Origin-Destination (OD) heterogeneity of route choice. To address this, we analyze one year of smartphone position data from over one million users in the Tokyo metropolitan area to extract high-resolution commuting trajectories. Our methodology is twofold: First, we develop algorithms to process raw position data, accurately extracting the commuting trajectory, transportation mode, and transfer stations. Second, by reinterpreting the Multinomial Logit (MNL) model through the canonical ensemble framework of statistical physics, we model route choice rationality as a temperature-dependent system. Our approach uniquely measures behavioral consistency in terms of rationality and preference stability over time, and distinguishes systematic from random heterogeneity. Our results reveal temporal stability in aggregate route choice behavior across the entire urban region throughout 2023. Also, we found heterogeneity dependent on the origin and destination (OD) pair. This variation is reflected as a bimodal split in the estimated route parameters, indicating that for certain attributes, commuters fall into two distinct groups with contrasting preference signs. We believe that our findings serve a basis for future urban route choice modeling by suggesting the importance of elabolating the model of transfer in railway.

Figures (12)

• Figure 1: Conceptual framework for this study. The methodology consists of four integrated components: (1) GPS point status classification and transfer station inference using raw trajectory data and station area analysis, where stroll status (regarded as stopping behavior in this study) indicates observable pauses during movement, such as traffic congestion or brief stops for car, bus, bike users, and transfers for railway users; (2) Commuting motif extraction to characterize diverse trajectory patterns based on stopping frequency; (3) Transport mode identification through mobility analysis (speed, acceleration, distance, etc.) and spatial analysis measuring proximity of moving points to railway infrastructure; and (4) Origin-Destination (OD) pair-based choice situation modeling using discrete choice analysis to measure route and station selection preferences.
• Figure 2: Activity state patterns and commuting motif types derived from GPS trajectory analysis. (a) Temporal distribution of five activity states (home, move, stay, stroll, work) over 24 hours. Clear diurnal patterns emerge: home activity dominates nighttime (0-6, 20-24 hours), movement peaks during morning (7-9) and evening (17-19) rush hours, and work activity sustains during business hours (9-17). (b) Distribution of commuting motif types by number of transfers. Direct routes without transfers (n=0) are most common, followed by single-transfer routes (n=1). Frequency decreases with increasing transfers; routes with three or more transfers (n$\geqq$3) are rare. The Other category includes patterns with stationary periods exceeding 40 minutes.
• Figure 3: Transportation mode identification process for railway and car commuters. Raw GPS trajectories (orange lines) are overlaid on the railway network (light blue lines) with 100-meter buffered zones (yellow). Dark blue segments indicate intersections between buffered zones and railway lines. Red stars mark identified trajectory points within these intersection areas. (a-b) Car commuter example: large distances between move status GPS points and railway network result in car mode classification. (c-d) Railway commuter example: short distances between move status GPS points and railway network result in railway mode classification.
• Figure 4: Aggregated trajectories for four recognized transportation modes. (a) Railway trajectories (red lines) closely follow the railway network (blue lines), demonstrating high spatial correlation with rail infrastructure. While some trajectories, such as those across the sea (Tokyo Bay), may appear discontinuous due to signal loss from tunnels or urban canyons, these instances are infrequent and do not compromise the overall dataset integrity for aggregate analysis. (b) Car or bus trajectories closely align the highway network (green lines) and urban road systems, instead of the railway network. Notably, the trajectories across Tokyo Bay align with the shape of the Tokyo Bay Aqua-Line. (c) Bicycle trajectories exhibit intermediate-range mobility patterns with moderate spatial coverage. (d) Walking trajectories display short-distance, localized movement patterns concentrated in urban areas. The distinct spatial characteristics of each mode validate the effectiveness of the transportation mode classification algorithm in this study.
• Figure 5: Model validation and sensitivity to scale parameter $\beta$. (a) Model validation showing predicted versus observed route choice probabilities. Black dots represent aggregated predictions with error bars indicating the 25th and 75th percentiles; grey dots show individual observations. The red dashed line indicates perfect prediction ($y=x$). Spearman $\rho = 0.72$. (b) Predicted route choice probability distribution and its relationship with scale parameter $\beta$. The red line shows model predictions with the estimated $\beta$ value; the light blue line shows the scenario with $\beta$ decreased by 1; darker blue lines show $\beta$ increased by 1 and 3, respectively. The dashed black line represents the observed probability distribution.