Identifying Contributory Factors to Heterogeneity in Driving Behavior: Clustering and Classification Approach

Previous research efforts using aerially collected trajectory-level data have confirmed the existence of inter-driver heterogeneity, where different car-following model (CFM) specifications and calibrated parameter sets are required to adequately capture drivers’ driving behavior. This research hypothesizes that there also exist clusters of drivers whose behavior is sufficiently similar to be considered a homogeneous group. To test this hypothesis, this study applies a 664-trip sample of trajectory-level data from the SHRP2 Naturalistic Driving Study to calibrate the Gipps, Intelligent Driver Model, and Wiedemann 99 CFMs. Using the calibrated parameter coefficients, this research provides evidence of the existence of homogeneous groups of driving behavior using the expectation maximization clustering algorithm. Four classification algorithms are then applied to classify the trip’s cluster ID according to driver demographics. Driver age, income, and marital status were most commonly identified as important classification attributes, while gender, work status, and living status appear less significant. The classification algorithms, which sought to classify a trip’s behavioral cluster ID by the driver-specific attributes, achieved the highest accuracy rate when predicting the desired velocity car-following parameter clusters. This effort illustrates that some drivers drive sufficiently alike to form a cluster of similar behavior; moreover, it was confirmed that driver-specific attributes can be utilized to classify drivers into these homogeneous driver groups.