Abstract
With the deployment of modern infrastructure for public transportation, several studies have analyzed movement patterns of people using smart card data and have characterized different areas. In this paper, we propose the “movement purpose hypothesis” that each movement occurs from two causes: where the person is and what the person wants to do at a given moment. We formulate this hypothesis to a synthesis model in which two network graphs generate a movement network graph. Then we develop two novel-embedding models to assess the hypothesis, and demonstrate that the models obtain a vector representation of a geospatial area using movement patterns of people from large-scale smart card data. We conducted an experiment using smart card data for a large network of railroads in the Kansai region of Japan. We obtained a vector representation of each railroad station and each purpose using the developed embedding models. Results show that network embedding methods are suitable for a large-scale movement of data, and the developed models perform better than existing embedding methods in the task of multi-label classification for train stations on the purpose of use data set. Our proposed models can contribute to the prediction of people flows by discovering underlying representations of geospatial areas from mobility data.
Introduction
Urban planning and marketing increasingly rely on understanding the functional roles of city districts. While demographic data provides a static snapshot, the true character of an area—whether it serves as a business hub, a residential suburb, or a student district—is dynamic and defined by human mobility. Traditional network embedding methods, often used in social media or linguistics, struggle with spatial data because movement is restricted by physical distance and transit infrastructure. This research addresses the challenge of creating mathematical representations (embeddings) of locations that respect physical constraints while capturing the underlying purpose of travel.
The Movement Purpose Hypothesis
The core of this research is the Movement Purpose Hypothesis: every trip is a result of two primary factors:
- Geography: Where a person is currently located and the physical constraints of the transit network.
- Purpose: What the person intends to achieve (e.g., commuting, shopping, school).
This hypothesis is formulated as a synthesis model where two distinct network graphs (geographic and purpose-based) generate the observed movement network graph.
Methodology: Novel Embedding Models
To evaluate the hypothesis, two novel embedding models were developed:
1. Concatenating Model
This model treats the geographic network and the purpose network as separate entities. It generates vector representations for each and then joins them together. This approach provides a baseline for how well independent features can describe movement.
2. Internally Dividing Model
This more sophisticated approach uses a learnable parameter ($\alpha$) to balance the influence of geography versus purpose. It assumes that the "People Flow" vector is a weighted sum of a "Geographic Constraint" vector and a "Purpose Proximity" vector. This allows the model to dynamically adjust the importance of each factor based on the data.
Both models adapt existing techniques like LINE (Large-scale Information Network Embedding) and PTE (Predictive Text Embedding) to handle the unique constraints of spatial data.
Experimental Setup and Datasets
The study focused on the Kansai region of Japan (including Osaka and Kyoto), utilizing three primary datasets:
- Smart Card Logs: Over 68 million "getting on and off" records from six railway companies, representing nearly 3.7 million unique users.
- Train Route Map: Used to define "Geographic Constraints" based on physical station connectivity.
- Person Trip Survey: A nationwide survey providing ground-truth labels for the purpose of visits at 599 stations (e.g., "commuting to work," "attending school," "on business").
Key Insights and Results
The models were evaluated on a multi-label classification task, predicting the primary reasons for visiting a station based on its embedding.
Performance Metrics
The proposed models outperformed standard benchmarks such as Word2Vec, GloVe, and DeepWalk. Specifically, the Internally Dividing Model achieved the highest accuracy in terms of KL Divergence and Mean Reciprocal Rank (MRR), validating the effectiveness of explicitly modeling the trade-off between distance and intent.
The 1.1x Rule
Through the Internally Dividing model, researchers estimated that the purpose of a visit is approximately 1.1 times more important than geographic distance in defining movement patterns within a large railroad network.
Functional Similarity
A significant finding was the ability to identify "Functional Similarity" over "Physical Proximity." The models successfully identified stations that are geographically distant but serve the same social function. For instance, business districts on opposite sides of the city shared similar "purpose" embeddings despite lacking direct physical connections. Visualization using t-SNE confirmed that while traditional methods cluster stations by train line, the proposed models cluster them by their urban role (e.g., grouping "school" stations together).
Conclusion and Future Work
This research demonstrates that network embedding methods are highly suitable for large-scale mobility data. By discovering the underlying representations of geospatial areas, these models can significantly improve people flow prediction and urban management.
Future work will explore:
- Integrating global graph structures (e.g., GraRep) to capture long-distance dependencies beyond two-hop proximity.
- Refining the use of labeled information to enhance the discovery of geospatial representations.
- Applying these embeddings to real-world urban planning to optimize transit and development.
Summary updated