POI Information Enhancement Using Crowdsourcing Vehicle Trace Data and Social Media Data: A Case Study of Gas Station

Wei Yang, Tinghua AiISPRS Int. J. Geo Inf 2018 · DOI

TL;DR

This paper introduces a novel method to enhance Point of Interest (POI) data, specifically for gas stations, by fusing sparse vehicle GPS trajectories with social media data. It addresses the limitations of using either data source alone by developing techniques to extract spatial information from GPS traces and semantic information from social media, ultimately creating a more comprehensive and up-to-date POI database.

What Problem Does This Paper Address?

Keeping Points of Interest (POIs) like gas stations accurate, complete, and up-to-date is crucial for many applications, from navigation to urban planning. Traditional methods like field surveys are costly and slow. While crowdsourced data offers a promising alternative, existing approaches have drawbacks: GPS trajectories lack semantic details (like names or reviews), and social media data can be outdated, inaccurate, or incomplete. This paper tackles the challenge of combining these disparate data sources to create a richer, more reliable POI database, focusing on gas stations as a case study.

Methods

The proposed method involves three main steps:

  1. Spatial Information Extraction from GPS Trajectories:
    • Refueling Stop Behavior Modeling: The authors first model the typical behavior of a vehicle stopping at a gas station (entering, refueling, leaving).
  2. Velocity Sequence Linear Clustering (VSLC): A novel algorithm is introduced to extract "refueling stop sub-trajectories" from sparse GPS data. This involves segmenting trajectories into "stop" and "move" states based on velocity and then clustering these segments. Crucially, it filters these stops to identify actual refueling events using the defined behavior model.
  3. Delaunay Triangulation: Collective refueling stop tracks are then used to construct a constrained Delaunay triangulation. By removing "long edges" (based on global and local thresholds), the algorithm extracts gas station polygons, which are then filtered using a "Gas Station Model" (considering direction and area) to identify precise gas station locations.
  4. Attribute and Semantic Information Extraction from Social Media:
    • Data Source: Social media data from Dianping (a popular Chinese web forum) is used.
  5. Text Mining and Tripartite Graph Model: A text mining approach is employed to extract attribute information (name, address, phone number) and dimensional sentiment (service, environment, products, subsidiary functions) from user reviews. This involves text preprocessing, identifying feature-opinion word pairs, calculating sentiment scores for sentences, and then aggregating these scores into dimension-specific scores for each gas station.
  6. POI Information Fusion and Enhancement:
    • Buffer Matching: The extracted spatial information (from GPS) and attribute/semantic information (from social media) are fused using a buffer matching method. A 50m buffer is created around GPS-derived POI points, and matching Dianping POIs within this buffer are merged. This step also handles cases of duplicate or outdated POIs.

Datasets

  • Vehicle GPS Trajectory Data: 15 days of taxi GPS trajectory data from 12,000 taxis in Beijing, China. This data includes taxi ID, latitude, longitude, speed, direction, and timestamp, with a sampling interval of 10-120 seconds (average 40 seconds).
  • Social Media Data: User-generated content (reviews, attribute information) for gas stations from Dianping (www.dazhongdianping.com).
  • Ground Truth Data: POI data for gas stations obtained from the National Geomatics Center of China, including location, name, address, type, and telephone number.

Key Insights and Ideas

  • Synergy of Sparse Trajectories and Social Media: The core insight is that neither sparse GPS data nor social media data alone is sufficient for high-quality POI enhancement. Their combination, however, unlocks richer information than either source individually.
  • Modeling Refueling Behavior is Key: Simply detecting "stops" in GPS data is insufficient. The paper's success hinges on accurately modeling the specific "refueling stop behavior" to distinguish it from other types of stops.
  • Delaunay Triangulation for Polygon Extraction: Using Delaunay triangulation on collective refueling tracks is an innovative way to derive the spatial footprint (polygon) of a gas station, rather than just a point.
  • Dimensional Sentiment Analysis: Moving beyond overall sentiment, the paper's approach to extracting sentiment for specific dimensions (service, product, etc.) provides much more actionable insights for users and managers.
  • Handling Data Sparsity and Noise: The VSLC algorithm is designed to be robust to sparse sampling and noise in GPS data, which is a common challenge in real-world trajectory datasets.

Key Results

  • Spatial Extraction: The VSLC algorithm demonstrated superior performance in extracting refueling stops compared to K-Medoids, DJ-Cluster, and CB-SMoT, achieving higher precision (0.83) and recall (0.76) on Dataset 1.
  • POI Enhancement: After fusion, 742 gas stations were enhanced. A qualitative evaluation against Google Earth imagery showed an accuracy of 93.5% (694 correct out of 742).
  • Attribute and Semantic Extraction: The method achieved high accuracy for attribute information (e.g., 98.6% for Name, 97.7% for Address) but lower integrity, as expected due to data incompleteness. Sentiment information (service, environment, products) was successfully extracted, providing insights into user perceptions.
  • Buffer-based Evaluation: Quantitative evaluation using buffer zones showed that 85.58% of extracted gas stations fell within a 30m precision range of the ground truth.

Limitations and Future Work

  • GPS Trace Limitations: The accuracy and completeness of GPS traces (sampling frequency, positioning errors) can lead to missed or incorrect extractions. Using longer time series or multiple types of vehicle traces could improve this.
  • Social Media Data Challenges:
    • Incompleteness and Outdatedness: Social media data can be incomplete, outdated, or contain duplicate entries, impacting the accuracy of attribute and semantic extraction.
  • Mining Complexity: Extracting sentiment information is complex and currently relies on manual dictionary construction. Future work should explore advanced techniques like Deep Learning for more accurate and automated sentiment mining.
  • Multimodal Data: Integrating other UGC sources like photos and videos could provide richer attribute data.
  • Detecting Dynamic Changes: The current method is limited in detecting dynamic changes like POI demolition, temporary maintenance, or relation changes. Future work should focus on integrating time information from comments and vehicle tracks, and fusing multi-web sources to capture these dynamics.
  • Data Quality Evaluation: Developing robust methods to evaluate and analyze the quality of data from different sources is crucial for effective multi-source data fusion.

Summary updated

BibTeX
@article{yang2018poi,
  author = {Wei Yang and Tinghua Ai},
  title = {{POI Information Enhancement Using Crowdsourcing Vehicle Trace Data and Social Media Data: A Case Study of Gas Station}},
  journal = {ISPRS Int. J. Geo Inf.},
  year = {2018},
  doi = {10.3390/ijgi7050178},
}
Figure