Data Analytics of Human Behaviors with Applications

Abstract

This dissertation presents a comprehensive analysis of human behaviors with an emphasis on practical methods that generate a positive societal impact with regards to driving safety and business management decisions. Human behaviors were explored through three major works: (1) automatic clustering of kinematic driving data patterns to enhance the understanding of drivers' normal maneuvering behaviors; (2) detecting distracted driving behaviors to enhance human driving safety through the fusion of kinematic data and driving models; and (3) modeling and predicting consumers' behaviors to assist airline management decisions. The major contributions are summarized as follows.

The first work introduces a new multi-model fusion and detection algorithm through effectively synthesizing three different state-space models, each of which represents a driver's typical kinematic motion pattern. The kinematic motion trajectory can be dynamically updated and predicted by using the Kalman filtering method. A normalized likelihood is further proposed for model fusion by weighing the contribution of each kinematic motion model. The results are then used to build two monitoring control charts. Specifically, both an Exponentially Weighted Moving Average (EWMA) control chart and a Cumulative Sum (CUSUM) control chart are developed to automatically detect and assess distracted driving behavior. This work is attractive in that it can be used to generate real-time warning signals during driving to avoid risky distracted driving behaviors.

The second major work aims to identify typical drivers’ normal maneuvering behaviors from a collected naturalistic driving dataset. This is done by improving the existing Hierarchical Dirichlet Process Hidden Semi-Markov Model (HDP-HSMM) inference method. In literature, a HDP prior can cause consistency issues, which harms the estimation of duration probabilities, transition probabilities, and total number of states estimated. Many solutions have been proposed to counter this issue; however, none have been from the perspective of HDP-HSMMs. This research presents a robust algorithm called rHDP-HSMM, which adds a state-merging procedure to the inference to avoid redundant states. As a result, the proposed rHDP-HSMM leads to more consistent states, faster convergence, and better parameter estimation than the existing HDP-HSMM method. The proposed rHDP-HSMM model is then used to enhance the understanding and inference of drivers’ maneuvering behaviors from a naturalistic dataset. The results can be served as a surrogated model to generate realistic simulation data for testing autonomous vehicles.

The third major work combines concepts from economics and constrained Gaussian Process (CGP) regression to probabilistically describe consumer purchasing behaviors. System-wide aggregation methods are often used to model customer buy-down/sell-up behaviors, however, the resulting estimates become uncharacteristic of the original markets of interest. This work addresses this issue by applying the multitask learning strategy to a CGP model, in which all fare groups' data are shared to estimate individual fare group demand simultaneously. Specifically, adding appropriate airline demand constraints facilitates an inter-group cross-learning capability. The proposed CGP model enables an accurate estimation of customers' purchasing behaviors in fare groups that have limited historical purchasing data. Buy-down and sell-up behaviors are then estimated by decomposing demand into three different states that consider which fare groups were available at the time of purchase. As such, this work offers both modeling contributions and a better understanding of customer purchasing behaviors in airline applications.

Collectively, the contributions from this dissertation have advanced the modeling of human behaviors across the transportation and business management fields.