Dimensionality Reduction to Reveal Urban Truck Driver Activity Patterns

This paper studies the activity profiles of truck drivers in urban areas. Finding repeating dynamical patterns is important in understanding freight behaviors, and aids freight-friendly planning. In the digital age, data on truck drivers is becoming more available with heterogeneous demographic and work profiles. By synthesizing such pervasive data and applying machine learning concepts, this paper proposes to identify signature travel activity patterns via dimensionality reduction. Based on driver survey data, truck drivers’ behaviors are represented as longitudinal activity sequences. Dimensionality reduction and activity reconstruction via principal components analysis (PCA), logistic PCA, and autoencoder were conducted to reveal fundamental activity features and approximate the underlying data-generating function. In the driver survey dataset, 243 truck drivers in Singapore reported their daily activities for 1,099 weekdays. This study found that PCA produced the most faithful reconstruction of drivers’ activities. When projecting the input data down from 2,592 to 82 dimensions, PCA explained 77% of variances with a reconstruction error of 0.99%. Logistic PCA is a useful extension of PCA to study the pattern of a single activity. It captures the variation of infrequent activities such as truck queuing, which PCA fails to reconstruct. Autoencoder was found to be more powerful than PCA in reconstructing activities – with 1% of original dimensions, it reconstructed the activities with an error rate of 1.24%. Moreover, when implemented as a variational autoencoder, autoencoder generated realistic-looking samples of driver activities. The top three most distinctive activity patterns of Singapore truck drivers are reported using PCA.


  • English

Media Info

Subject/Index Terms

Filing Info

  • Accession Number: 01661241
  • Record Type: Publication
  • Report/Paper Numbers: 18-03813
  • Files: TRIS, TRB, ATRI
  • Created Date: Jan 8 2018 10:57AM