Evaluating Origin-Destination Matrices Obtained from CDR Data

Marco Mamei, Nicola Bicocchi, Marco Lippi, Stefano Mariani, Franco ZambonelliSensors 2019 · DOI

TL;DR

This paper tackles the challenge of accurately estimating urban mobility patterns by developing and evaluating methods to generate Origin-Destination (OD) matrices from Call Detail Records (CDRs). The research highlights the trade-offs between time-based and routine-based approaches, offering a framework for understanding and validating these estimations against real-world data.

What Problem Does This Paper Address?

Understanding and modeling urban mobility is crucial for developing smart cities, optimizing transportation systems, and informing urban planning. A key component of this is the Origin-Destination (OD) matrix, which quantifies travel demand between different zones. Traditionally, this data comes from expensive and time-consuming surveys. The advent of mobile phone technology has opened up new possibilities for collecting vast amounts of location data through Call Detail Records (CDRs). However, CDRs provide only sporadic location samples, and extracting accurate OD matrices from this noisy data is a significant challenge. Existing methods often rely on either directly analyzing sequences of CDRs (time-based) or inferring routine movements like home-work commutes (routine-based). This paper aims to provide a comprehensive evaluation of these approaches, detailing their implementation and validating their results.

Methods

The paper proposes a methodology for estimating OD matrices from CDR data, focusing on two main types:

  • Time-based OD matrices (tOD): These matrices are derived by directly analyzing sequences of CDRs within specific time windows. Each CDR pair is interpreted as a trip between the cell towers the user was connected to.
  • Routine-based OD matrices (rOD): This approach specifically targets routine movements, with a strong emphasis on the home-work commute (hwOD). It involves identifying home and work locations for users based on their CDR patterns and then estimating commuting trips. A novel mechanism is introduced to estimate departure and arrival times, addressing data sparsity issues in existing methods.

Both approaches involve:

  1. Trip Identification: Inferring individual trips from CDR sequences or routine patterns.
  2. Data Aggregation: Aggregating individual trips into OD matrices, using either a "starting time rule" or an "ending time rule" for trip assignment to time intervals.
  3. Scaling: Adjusting the raw OD counts to match actual population numbers and market penetration rates of mobile phone users.
  4. Road Assignment: Projecting the OD flows onto the road network using an incremental traffic assignment (ITA) algorithm, which iteratively updates travel costs based on assigned traffic volume. This step uses the A* algorithm for route finding and incorporates road capacity information from OpenStreetMap.

Datasets

  • Call Detail Records (CDRs): Obtained from an Italian telecom operator, covering three regions (Piemonte, Lombardia, and Emilia Romagna) with approximately 20 million people. Data spans several months.
  • ISTAT Census Data (2011): Used for ground-truth information on residential distribution and commuting patterns (origin-destination matrix of journeys for work or study).
  • Google Maps: Used for visual comparison of road-level traffic estimations.

Key Insights and Ideas

  • Routine-based matrices (hwOD) show higher correlation with census data: This suggests that inferring structured routines, like home-work commutes, provides a more robust basis for OD estimation compared to simply aggregating all observed movements.
  • Time-based matrices capture broader mobility but are less precise for specific routines: While time-based matrices can reflect all types of movements, their correlation with structured commute data is lower, likely due to capturing daily variations and non-routine trips.
  • Home and workplace identification is crucial but challenging: The accuracy of routine-based OD matrices heavily relies on correctly identifying users' home and work locations, which can be biased by data sparsity and the spatial resolution of CDRs.
  • Road assignment is sensitive to spatial uncertainty: Randomly assigning home/work locations within a municipality can lead to significant errors in travel time estimations, especially in complex terrains like mountainous regions.
  • OD matrix information is dominant over network layout for traffic prediction: Experiments show that the OD matrix itself contains significantly more information about traffic patterns than the road network layout alone.

Key Results

  • Routine-based OD matrices (hwOD) achieved an average R² of 0.8 when compared to ISTAT census data for home-work commutes.
  • Time-based OD matrices (tOD) achieved an average R² of 0.5 when compared to ISTAT data, indicating they capture broader mobility but are less precise for specific routines.
  • Home distribution correlation with census data varied: R² values ranged from 0.46 (Piemonte) to 0.64 (Emilia Romagna).
  • Comparison with Google Maps traffic: Preliminary analysis showed alignment between the paper's road-level traffic estimations and Google Maps, though differences were observed, potentially due to how traffic is displayed.
  • Random OD matrix comparison: Demonstrated that the OD matrix itself explains over 90% of the traffic information, with the transportation network layout explaining less than 10%.

Limitations and Future Work

  • Data Sparsity: The limited number of CDRs per user per day poses a challenge for accurately identifying trips and routines.
  • Spatial Resolution: CDRs provide location at the cell tower level, which can be a coarse approximation for precise origin/destination points, especially in rural areas.
  • Simplistic Assumptions: The routine-based model makes simplified assumptions about commute probabilities and timing.
  • Future Work:
    • Improve analysis by incorporating data from other sources like traffic cameras and sensors.
  • Integrate time-based and routine-based OD matrices for a more comprehensive understanding of mobility.
  • Refine road traffic assignment models to account for individual travel preferences.
  • Develop methods to highlight non-routine travels by subtracting routine OD matrices from time-based ones.
  • Build an automatic traffic bulletin system based on these insights.

Summary updated

BibTeX
@article{mamei2019evaluating,
  author = {Marco Mamei and Nicola Bicocchi and Marco Lippi and Stefano Mariani and Franco Zambonelli},
  title = {{Evaluating Origin-Destination Matrices Obtained from CDR Data}},
  journal = {Sensors},
  year = {2019},
  doi = {10.3390/s19204470},
}

Loading related content...

Figure