Identifying the Multimodal Hierarchy of Public Transit Systems Using Itinerary Data

As urban mobility integrates traditional and emerging modes, public transit systems are becoming increasingly complex. Some modes complement each other, while others compete, influencing users’multimodal itineraries. To provide a clear, high-level understanding of these interactions, we introduce the concept of a macroscopic multimodal hierarchy. In this framework, trips follow an"ascending-descending"order, starting and ending with lower hierarchical modes (e.g., walking) that offer high accessibility, while utilizing higher modes (e.g., subways) for greater efficiency. We propose a methodology to identify the multimodal hierarchy of a city using multimodal smart card itinerary data and demonstrate its application with actual data collected from Seoul and the surrounding metropolitan area in South Korea.

💡 Research Summary

**
The paper introduces a novel framework called the macroscopic multimodal hierarchy to capture the high‑level interactions among various transportation modes in modern urban mobility systems. The authors posit that most multimodal trips follow an “ascending‑descending” pattern: they start with low‑hierarchy modes (e.g., walking or cycling) that provide high accessibility, move up to higher‑hierarchy modes (e.g., subways or express buses) for speed and capacity, and finally descend back to low‑hierarchy modes for the final leg. To operationalize this concept, the study develops a data‑driven methodology that extracts hierarchical structures directly from smart‑card itinerary records.

First, raw transaction logs (tap‑in/out timestamps, station IDs, and line identifiers) are cleaned and linked to reconstruct individual itineraries. Each segment of an itinerary is labeled with its transport mode, producing a sequence of mode labels per trip. From these sequences, a mode‑to‑mode transition matrix is built, where each cell records the frequency and average travel time of a specific transition. The authors introduce asymmetric cost weights: upward transitions (low‑to‑high hierarchy) receive lower penalties, while downward transitions (high‑to‑low) are penalized more heavily, reflecting the hypothesized efficiency gain when moving upward.

Second, a rank‑based optimization model is employed to infer the most plausible ordering of modes. The model treats the hierarchy as a permutation of modes and minimizes a loss function that captures the mismatch between observed transition probabilities and the idealized ascending‑descending flow. The number of hierarchy levels (k) is treated as a hyper‑parameter; models with k ranging from three to seven are evaluated using Akaike Information Criterion and cross‑validation to select the best‑fitting structure.

The methodology is applied to a massive dataset from Seoul and the surrounding Gyeonggi province, comprising over 20 million smart‑card transactions collected throughout 2022. The analysis reveals a five‑level hierarchy that best explains the observed travel patterns: (1) walking/cycling, (2) local buses, (3) regional buses, (4) subway/metro, and (5) express/airport buses. During peak commuting hours, the dominant transitions are from level 3 to level 4, indicating a strong reliance on subways for the core of the trip, while late‑night periods show a simplified pattern of level 2 to level 1, reflecting reduced demand for high‑speed services.

The authors discuss several policy implications. By recognizing the hierarchical nature of trips, transit agencies can design targeted transfer incentives—such as fare discounts for walking‑to‑subway connections—that encourage smoother ascents and descents, thereby improving overall network efficiency. Infrastructure investments can be prioritized to strengthen the low‑hierarchy “first‑ and last‑mile” links (e.g., expanded sidewalks, protected bike lanes) that feed into higher‑hierarchy services. Real‑time monitoring of hierarchical transition flows can also support dynamic congestion management and service adjustments.

Limitations are acknowledged: the smart‑card data omit non‑carded modes such as taxis, ride‑hailing, and private car trips, and the model does not directly incorporate trip purpose or individual user preferences. Future work is suggested to integrate GPS‑based personal travel logs, apply deep‑learning techniques to capture non‑linear mode interactions, and conduct cross‑city comparative studies to explore how cultural and geographic factors shape multimodal hierarchies.

In sum, the paper provides a rigorous, scalable approach to uncovering the latent hierarchical organization of multimodal transit systems, offering both a theoretical lens and a practical tool for planners seeking to orchestrate more coherent and efficient urban mobility networks.