What Kind of Travellers Are Using Carsharing in Beijing? A Study Based on Selective Ensemble Learning

Wei Luo, Yi Wang, Pengpeng Jiao, Zehao Wang, Pengfei ZhaoSustainability 2022 · DOI

TL;DR

This paper uses a novel selective ensemble learning approach to analyze carsharing user data in Beijing, identifying four distinct traveler profiles. Understanding these profiles is crucial for optimizing carsharing services, urban planning, and promoting sustainable transportation.

What Problem Does This Paper Address?

Carsharing is a growing urban mobility solution, but its effectiveness is hampered by a lack of understanding regarding who uses it and why. Traditional transportation modes have well-defined user groups, but carsharing's unique characteristics (e.g., on-demand, self-service) create a more complex user base. This paper aims to bridge this knowledge gap by identifying and characterizing distinct carsharing traveler types in Beijing, moving beyond generic user profiles to a more nuanced understanding of their travel behaviors, personal attributes, and environmental contexts. This detailed understanding is vital for carsharing operators to tailor services, for urban planners to integrate carsharing effectively into the broader transportation ecosystem, and for policymakers to promote sustainable urban mobility.

Methods

The core of this research lies in the application of a selective ensemble learning model to analyze carsharing travel data. The authors first construct a "unified view of travel" by integrating data from three sources: user registration systems, order logs, and GPS-based vehicle monitoring. This unified view is then broken down into three distinct "views": personal characteristics, environmental characteristics, and behavioral characteristics.

For each view, unsupervised learning algorithms (specifically, a clustering algorithm based on minimizing within-cluster variance, likely a variant of K-means) are used to generate initial "base classifiers." The novelty comes in the selective ensemble learning stage. Instead of simply combining all base classifiers, this method intelligently selects the most informative ones. This involves:

  1. Constructing a Base Classifier Collective: Generating multiple clustering results from different subsets of the data or feature sets.
  2. Evaluating Base Classifiers: Using a novel evaluation function that considers both cluster tightness (average similarity within clusters) and inter-cluster separation (using a similarity metric and a weighting factor). This helps identify high-quality base classifiers.
  3. Base Classifier Integration: Merging the selected base classifiers using a consistency function that evaluates the overlap and representativeness between clusters from different views. This process aims to achieve a more accurate and robust classification of carsharing travel patterns.

The model's performance was validated by comparing it against standard K-means clustering using normalized mutual information (NMI) as an evaluation metric.

Datasets

The study utilizes a comprehensive dataset from a carsharing company in Beijing, collected over one month. This dataset comprises:

  • User Order Data: Includes origin/destination station IDs, reservation time, user age, and sex.
  • Vehicle GPS Data: Provides time, vehicle ID, longitude, latitude, charging status, electric quantity, and total mileage, updated every 10 seconds.
  • Vehicle Information Data: Contains vehicle ID, license plate number, and identification number (anonymized for privacy).
  • Station Information Data: Includes station names, IDs, and addresses.

Additionally, weather conditions and air quality data were obtained from the Chinese meteorological network to enrich the environmental characteristics.

Key Insights and Ideas

  • Beyond Generic Users: Carsharing isn't a one-size-fits-all service. The research clearly demonstrates that users can be segmented into distinct groups with unique travel needs and preferences.
  • The Power of Ensemble Learning: By combining multiple learning models and intelligently selecting the best ones, the selective ensemble learning approach significantly outperforms traditional K-means clustering, especially with larger datasets. This highlights the benefit of leveraging diverse perspectives for complex classification tasks.
  • Interplay of Factors: The study emphasizes that understanding carsharing travel requires considering personal, environmental, and behavioral factors in conjunction, rather than in isolation.
  • Context Matters: Environmental factors like weather and air quality, while not always dominant, can influence travel choices, particularly for leisure-oriented trips.
  • Carsharing as a Complement, Not Just a Substitute: The analysis suggests that carsharing can complement public transport and other modes, rather than solely competing with them, especially for specific travel needs.

Key Results

The selective ensemble learning model successfully identified four distinct types of carsharing travel:

  1. Long Distance for Leisure and Entertainment (LDLE): Primarily weekend travelers, often using carsharing for recreation, with long travel durations and distances, typically in good weather.
  2. Medium and Short Distances for Business and Commuting (MSDBC): Weekday commuters, using carsharing for shorter, inflexible trips between business and commercial areas, less influenced by external environmental conditions.
  3. Mixed Category of Medium and Short Distances for Business and Residence (MSDBR): Travelers using carsharing for short to medium distances, potentially influenced by bad weather (like rain), suggesting a comfort-driven choice for commuting and shopping.
  4. Mixed Category of Long Distance for Business and Residence (LDBR): Weekday travelers (Mondays/Tuesdays) for medium to long distances, primarily for commercial purposes, with destinations often in retail and commercial areas.

The study also found that men are the dominant users of carsharing in Beijing, a finding that may differ from studies in other cities.

Limitations and Future Work

The authors acknowledge several limitations:

  • Limited Influencing Factors: The study considered only 11 travel characteristics, with personal characteristics limited to sex and age. Future work should incorporate more complex personal factors like occupation, income, family structure, and even driving experience or accident history.
  • Further Competition Analysis: A deeper investigation into the competition and complementarity between carsharing and other urban travel modes is suggested.
  • Broader Context: Exploring the influencing factors and competition processes of carsharing in a wider range of urban contexts beyond Beijing would be beneficial.

The authors suggest that future research should consider additional personal, environmental, and behavioral characteristics, as well as delve deeper into the interplay between carsharing and other transportation systems for more comprehensive urban mobility planning.

Summary updated

Abstract

As a new urban travel mode, carsharing is significantly different from private cars, buses and other travel modes. Therefore, clarifying the typical characteristics of carsharing, such as individual users’ attributes, travel environment and travel behaviour, is conducive to accurately grasping the development of carsharing. In this study, a selective ensemble learning model is established to analyse typical travel characteristics of carsharing. Firstly, personal characteristics, environmental characteristics and behavioural characteristics were obtained through integrating order data, global positioning system data and station information. Then, based on a consolidated view of carsharing, different types of carsharing travel characteristics were distinguished using selective ensemble learning. Lastly, all kinds of carsharing travel are described in detail. It was identified through this research that carsharing travel can be divided into four kinds: long distance for leisure and entertainment, medium and short distances for business and commuting, a mixed category of medium and short distances for business and residence, and a mixed category of long distance for business and residence. This study can provide a theoretical reference and practical basis for precise planning and design and the scientific operation of carsharing.
BibTeX
@inproceedings{luo2022what,
  author = {Wei Luo and Yi Wang and Pengpeng Jiao and Zehao Wang and Pengfei Zhao},
  title = {{What Kind of Travellers Are Using Carsharing in Beijing? A Study Based on Selective Ensemble Learning}},
  booktitle = {Sustainability},
  year = {2022},
  doi = {10.3390/su14010540},
}
Figure