Abstract
This paper introduces a novel framework for inferring sociodemographic attributes—such as age, income, and household structure—from passive mobility data. By representing individual travel patterns as directed graphs and employing a multitask learning (MTL) architecture, the researchers significantly improve prediction accuracy and model reliability. The approach leverages "higher-order" behavioral features, including motifs and co-travel statistics, demonstrating particular resilience in data-scarce scenarios and during the significant travel behavior shifts following the COVID-19 pandemic.
Introduction
Modern transportation planning increasingly relies on "passive" mobility data, such as GPS traces from smartphones or location pings from mobile applications. While these datasets offer massive scale at low cost, they suffer from a critical "blind spot": they record where and when movement occurs, but lack information on who is moving.
Understanding the demographics of travelers is essential for addressing equity—ensuring transit reaches underserved communities—and for predicting how different groups respond to policy changes. However, predicting demographics from movement is notoriously difficult because the relationship between behavior and identity is often weak, noisy, and highly context-dependent. This research bridges this gap by introducing sophisticated graph-based features and robust machine learning techniques to extract demographic signals from raw mobility data.
Methodology
The authors approach the prediction problem through three distinct technical lenses:
1. Mobility Graphs
Instead of relying on simple aggregate metrics like total distance traveled, the framework encodes an individual's daily activity as a directed graph. In this representation, nodes represent activity purposes (e.g., Home, Work, Social) and edges represent the trips between them. This structure allows for the extraction of "higher-order" features that capture the nuance of human behavior.
Key features extracted include:
- Motifs: Recurring daily subgraphs, such as simple "out-and-back" trips versus complex multi-stop loops.
- Clustering Characteristics: How interconnected an individual's destinations are.
- Co-travel Statistics: Whether a person typically travels alone or with household members, which serves as a powerful proxy for household structure.
2. Multitask Learning (MTL)
Rather than training isolated models for age, gender, and income, the researchers utilize a "shared-trunk" neural network. This model learns a joint representation of mobility that feeds into four specific task "heads." This architecture exploits natural correlations between demographic attributes; for instance, travel patterns signaling the presence of children are often intrinsically linked to specific age brackets.
3. Uncertainty Quantification
The study emphasizes model calibration to ensure reliability. Using tools like reliability diagrams and Expected Calibration Error (ECE), the authors ensure that the model’s confidence scores accurately reflect its predictive accuracy, preventing the overconfidence common in deep learning models.
Datasets
The study utilizes the Puget Sound Regional Council (PSRC) Household Travel Survey (HTS) from the Seattle region, analyzing three distinct waves: 2017, 2019, and 2023. This dataset is uniquely valuable as it provides "ground truth" demographics paired with detailed trip diaries.
The 2023 wave is particularly significant as it represents a major "distribution shift." Because travel behavior changed fundamentally after the COVID-19 pandemic, this wave provides a rigorous test of how well models trained on historical data can generalize to a new social reality.
Key Results and Insights
- Accuracy and Feature Power: The inclusion of higher-order graph descriptors (motifs and co-travel) consistently improved out-of-sample accuracy and AUROC across all demographic categories. Co-travel features emerged as the strongest predictors for household structure and age.
- Sample Efficiency: The Multitask (MT) variant proved significantly more robust than single-task models when training data was scarce. In experiments where the training set was reduced to just 0.1% or 1% of its original size, the MT model maintained substantially higher performance.
- Resilience to Temporal Shifts: Models trained on pre-2020 data naturally struggled with 2023 data. However, the graph-based features and MTL approach were more resilient to these shifts than standard spatiotemporal baselines, suggesting that the underlying "grammar" of mobility is more stable than simple metrics like speed or distance.
- Behavioral Grounding: Moving toward semantic features (purpose, motifs) provides the interpretability necessary for urban planners to trust and act upon model outputs.
Limitations and Future Work
While the framework shows great promise, several challenges remain. The current study uses "clean" trip diaries where purposes and modes are pre-labeled. Real-world deployment using raw GPS pings would require an initial "imputation" step to guess trip purposes, which may introduce noise. Additionally, the models were tested exclusively in the Seattle area; future research must evaluate "cross-city" portability to determine if models trained in one urban environment can successfully generalize to others with different social norms and urban forms.
Summary updated