Forecasting Public Transit Use by Crowdsensing and Semantic Trajectory Mining: Case Studies

Ningyu Zhang, Huajun Chen, Xi Chen, Jiaoyan Chen

TL;DR

This paper proposes a novel approach to forecast public transit usage by combining crowdsensing data with semantic trajectory mining. By enriching user movement data with contextual information like Points of Interest (POIs) and weather, the authors achieve improved predictions for both individual passenger boarding choices and overall passenger flow.

What Problem Does This Paper Address?

As smart cities grow, public transit systems face increasing pressure. Accurately forecasting passenger demand is crucial for transit providers and city planners to optimize routes, manage resources, and improve the commuter experience. Traditional methods often rely solely on historical travel patterns, which can be insufficient as they don't account for external factors influencing travel decisions. The authors aim to bridge this gap by leveraging the rich, often underutilized, data generated by users' mobile devices and public transit infrastructure.

Methods

The core of the proposed method involves two main stages: feature engineering and model building.

  1. Data Preprocessing and Semantic Enrichment:
    • Dirty Data Cleaning: The raw bus card transaction data undergoes cleaning to remove anomalies and inconsistencies, such as terminal IDs associated with multiple bus lines or multiple records for the same card on the same line at the same time.
  2. Semantic Trajectory Mining: User trajectories are enriched with semantic information. This involves mapping locations to POIs (e.g., "Coffee Shop") and neighborhood functions (e.g., "Education," "Business"). This transforms raw GPS points into meaningful sequences of activities or locations.
  3. Feature Engineering: A comprehensive set of features is constructed, including:
  4. Passenger-specific features: Number of records, hours, days, weeks, and distinct terminals visited over various time windows (1, 3, 7, 28, 70, 126 days).
  5. Bus line features: Line name.
  6. Time features: Weekday, weekend, holiday, rush hour.
  7. Weather features: Categorized temperature and weather conditions.
  8. Semantic user behavior features: Derived from the semantic trajectories.
  9. Bus card issuing location: City of issuance.
  10. Prediction Model Building:
    • XGBoost: The authors employ XGBoost, a powerful gradient boosting algorithm, for both classification (predicting if a passenger will take a specific bus) and regression (predicting passenger flow). XGBoost is chosen for its scalability, speed, and effectiveness in handling complex feature interactions.
  11. Hybrid Passenger Flow Model: For passenger flow forecasting, they recognize the distinct patterns of "frequent" and "occasional" passengers. They build separate models for each group and then combine their predictions using learned weights (α and β) that adapt to weekday/weekend and holiday variations.

Datasets

The study utilizes real-world datasets from Guangzhou, China:

  • Bus Card Transaction Data: Provided by the Ridge Nantong Limited bus company, covering a period of several months (August 2014 - December 2014). This data includes bus card ID, terminal ID (bus stop ID), and travel time.
  • Public Bus Information: Details about bus routes.
  • Weather Data: Guangzhou weather information.
  • Point of Interest (POI) Data: From Baidu Map API, providing names and types of POIs.
  • District Function Data: Categorizing urban areas (e.g., Community, Business, Park).

The data was split into training (December 2014) and testing (January 1-7, 2015) sets.

Key Insights and Ideas

  • Semantic Enrichment is Crucial: Simply using geographic or temporal data is insufficient. Incorporating semantic labels (like "Education" or "Business" for a location) significantly improves prediction accuracy by capturing the why behind travel.
  • Context Matters: External factors like weather and time of day (weekday vs. weekend) have a substantial impact on transit choices, and accounting for them is vital.
  • User Behavior Heterogeneity: Passengers are not monolithic. Distinguishing between frequent and occasional riders is key to accurately forecasting overall passenger flow.
  • XGBoost's Power: The study demonstrates XGBoost's superior performance over traditional methods like logistic regression, GBDTs, and Random Forests for boarding choice prediction, and over ARMA, ANN, and linear regression for passenger flow forecasting.

Key Results

  • Passenger Boarding Choice Forecasting: The proposed method, using all features (time, weather, semantic), achieved the best results, outperforming baselines. Adding weather and semantic features individually led to significant improvements over using only time features. XGBoost consistently outperformed logistic regression, GBDT, and Random Forest.
  • Passenger Flow Forecasting: Similarly, the "all features" approach with XGBoost yielded the lowest Root Mean Square Error (RMSE), indicating superior accuracy in predicting passenger flow compared to linear regression, ANN, and ARMA.

Limitations and Future Work

  • Data Sparsity and Missing Information: A major limitation is the partially missing data, particularly the lack of explicit "get off" timestamps for many passengers. This hinders a deeper understanding of individual passenger behavior.
  • Limited Open Data Availability: The study notes the general challenge of limited open data for comprehensive analysis.
  • Future Work:
    • Visual Analytics: Integrating visual analytics functions into a demonstration system to provide more flexible decision-making insights.
  • New Prediction Models: Investigating prediction models that leverage data from similar historical circumstances by understanding underlying data semantics.
  • Deeper Behavior Mining: Aiming to mine passenger behaviors more deeply, potentially by addressing the missing "get off" data or by inferring it.
  • Adaptability: Enhancing the approach's adaptability to real-world circumstances.

Summary updated

Abstract

With the growing development of smart cities, public transit forecasting has begun to attract significant attention. In this paper, we propose an approach for forecasting passenger boarding choices and public transit passenger flow. Our prediction model is based on mining common user behaviors for semantic trajectories and enriching features using knowledge from geographic and weather data. All the experimental data comes from the Ridge Nantong Limited bus company and Alibaba platform which is also open to the public. We evaluate our approach using various data sources, including point of interest (POI), weather condition, and public bus information in Guangzhou to demonstrate its effectiveness. Experimental results show that our proposal performs better than baselines in the prediction of passenger boarding choices and public transit passenger flow.
BibTeX
@article{zhang2016forecasting,
  author = {Ningyu Zhang and Huajun Chen and Xi Chen and Jiaoyan Chen},
  title = {{Forecasting Public Transit Use by Crowdsensing and Semantic Trajectory Mining: Case Studies}},
  journal = {ISPRS Int. J. Geo Inf.},
  year = {2016},
  doi = {10.3390/ijgi5100180},
}
Figure