Feature Selection in a Personalized Refueling Recommendation System for Drivers

Techniques for feature selection are evaluated in the context of a personalized recommendation system for drivers. The system uses a content-based filtering approach to provide personalized recommendations for the best available fuel station. The candidate fuel stations are characterized by their price, detour from the driver’s route, and brand. Recommendations are formed by combining these attributes with driver preferences determined from the driver’s refueling history. Limited training data and a relatively large number of possible brands can severely limit the learning speed of the recommendation engine. Several techniques for feature selection are applied to the brand features and their performance is evaluated using a simulation. A method based on the fraction of the time the user chooses a certain brand is shown to perform best in the most realistic simulation scenario, giving ultimate mean precision of about 80% for drivers who have no brand preference and about 85% for drivers who have a strong brand preference.

  • Supplemental Notes:
    • This paper was sponsored by TRB committee AND20 Standing Committee on User Information Systems.
  • Corporate Authors:

    Transportation Research Board

    500 Fifth Street, NW
    Washington, DC  United States  20001
  • Authors:
    • Thompson, Joshua
    • Zhao, Yunjie
    • Jang, Dongwook
    • Giurgiu, Gavril
  • Conference:
  • Date: 2016


  • English

Media Info

  • Media Type: Digital/other
  • Features: Figures; References;
  • Pagination: 12p
  • Monograph Title: TRB 95th Annual Meeting Compendium of Papers

Subject/Index Terms

Filing Info

  • Accession Number: 01590310
  • Record Type: Publication
  • Report/Paper Numbers: 16-2034
  • Files: TRIS, TRB, ATRI
  • Created Date: Jan 12 2016 4:53PM