Complete Trip: A Linked Multimodal Human Mobility Dataset

Connecting the Dots of Human Movement

Can a researcher tell if a person missed a bus because the schedule was bad or because they lived too far from the stop? Most mobility data fails this test. They might show a person arrived at a station, but they miss the "first-mile" walk that preceded it. This fragmentation makes it difficult to understand the full complexity of multimodal travel—the chain of events where a person walks, takes a train, and then catches a bus.

Understanding how people move is vital for designing bus routes or modeling disease spread. Historically, researchers have relied on fragmented data. They might use GPS traces that show a dot moving on a map. Alternatively, they might use aggregate reports showing how many people visited a park. Rarely do they have both. This creates a massive gap in our understanding of how different modes of transport actually connect.

Existing mobility datasets typically capture only isolated aspects of travel behavior. Some offer high-resolution trajectories but lack info on whether a person was driving or cycling. Others provide large-scale aggregate indicators but lose the individual journey entirely. Currently, no publicly available dataset combines multimodal travel, linked trips, network-based route representations, and population-level inference.

The fragmentation of modern mobility data

Current approaches to studying movement often force a trade-off between detail and scale. High-fidelity trajectory datasets, such as GeoLife, capture granular movement. However, they generally lack the multimodal context needed to understand complex commutes. Conversely, aggregate products like Google Mobility Reports provide vast spatial coverage. Yet, they only offer broad indicators rather than the linked individual journeys necessary to see how one person navigates a city.

Even survey-based datasets provide rich behavioral information. However, they are often geographically constrained. They cannot be easily scaled to represent an entire population. As shown in [Table 1], existing datasets tend to excel in one area while failing in others. This fragmentation makes it difficult for urban planners to perform "multimodal accessibility assessments." This is the process of measuring how easily people can reach essential services using various transport modes. Without linked data, they cannot reconstruct the full "trip chain" from a person's front door to their office.

Reconstructing the complete journey

To solve this, the authors introduce "Complete Trip." This framework transforms raw, passive smartphone location-based services (LBS) data into a structured, linked representation of human mobility. The process follows a four-stage workflow illustrated in :

1. Trip Identification: The pipeline begins with raw "pings" (timestamped geographic coordinates). It uses the FHWA NextGen NHTS framework to segment these observations into individual trips. It identifies activity anchors to separate movement from periods of rest.
2. Mode Imputation: Once trips are identified, the system determines how they occurred. The authors use a supervised Random Forest (RF) classifier to assign each trip to one of four modes: car, bus, rail, or active transportation (walking/biking). This classifier uses trip behavior and the surrounding transportation network context.
3. Route Reconstruction: The pipeline maps trips onto digital networks. Car and active trips are matched to OpenStreetMap (OSM) networks using a specialized Hidden Markov Model called Sup-HMM. Transit trips are reconstructed using the General Transit Feed Specification (GTFS) to include specific access and egress stops.
4. Trip Linking: Finally, the system links adjacent trip segments that belong to the same "travel episode." By linking these segments, the dataset can represent a single journey involving multiple modes. This includes a walk followed by a bus ride, as seen in .

Validating the reconstructed reality

The authors validate the dataset by comparing its outputs against official, independent benchmarks. In the Utah six-county study area, the researchers report that reconstructed car travel volumes closely follow official UDOT traffic counts. This reproduces the major temporal variations observed in the real world [Figure 4b].

A significant achievement of the paper is the application of "population expansion weighting." Smartphone data is often biased toward certain demographics. To fix this, the authors apply a two-stage weighting framework to align the data with census totals. The authors report that after applying these weights, the reconstructed transit trips show a Pearson correlation of $\rho = 0.96$ with official UTA ridership records [Figure 4f]. This near-perfect correlation shows the dataset effectively represents actual population-level transit demand.

The scale of the processed data is substantial. From an initial pool of over 6.3 billion raw pings, the authors reconstructed over 28 million trip records and 28 million linked journeys [Table 3].

Limitations of passive sensing

Despite the sophisticated reconstruction, the authors note several inherent limitations. First, the dataset relies on passive LBS data. This means it cannot account for populations that generate no digital footprint. This includes young children or individuals without smartphones. Weighting can mitigate demographic bias, but it cannot "create" data for populations that are entirely absent from the raw signal.

Second, the data uses privacy-preserving transformations that limit precision. To protect identities, the authors generalize trip origins and destinations to Geohash-6 grid cells (roughly 1km x 1km areas). They also abstract timestamps into 30-minute intervals. This allows for robust population-level analysis. However, it means the dataset cannot be used for micro-scale studies requiring exact street addresses or second-by-second timing.

Finally, the authors note that "residual sampling bias" remains. Even with rigorous statistical calibration, the dataset is still a sample of a sample. Practitioners should interpret the results as estimates rather than absolute ground truths.

A new tool for urban intelligence

The Complete Trip dataset is ready for high-level urban science and transportation research. It is well-suited for advanced applications like "transit desert" analysis. This helps identify areas where people want to use transit but face structural barriers, such as long walking distances to the nearest stop . It can also analyze tourism patterns by linking hotel stays to subsequent local activities .

For researchers looking to implement this, the framework is geographically transferable. Code is reportedly available via the project's GitHub repository. If your goal is to model city-wide demand or evaluate transit equity, this dataset provides the linked, multimodal connectivity that previous generations of mobility data lacked.

Figures from the paper