The goal of this project is increasing road safety by understanding and predicting driver engagement via the analysis of 4 Petabyte of SHRP 2 naturalistic driving data.  In this project, we first conduct a thorough data pre-processing, including reviewing variables, feature selection, data cleaning, de-duplication, filtering, format conversion, data integration and fusion, and data normalization, to ensure the data quality for predictive analytics.  We develop heterogeneous network mining and sophiscated statistical analysis to construct a global picture of driving behavior analysis focusing on driver engagement.  The temporal heterogeneous network mining are developed to extract the frequent patterns.  Novel representation is proposed to preserve the temporal aspects of the heterogeneous networks.  Constrained interval graphs are introduced and algorithms are developed to extract a complete set of frequent temporal patterns with isomorphism.  These frequent patterns are tested in the prediction model.  Evolution of temporal pattern in heterogeneous network, such as growing, diminishing, merging, or spliting, are identified  to understand the changes of patterns in sensor variables and their associations.  These patterns are helpful in the explanations of the prediction. We integrate a distributed and asynchronous computing infrastructure to support the distributed searching and asynchronous parallel graph computing.   Aggressive driving behavior are identified and analzyed with t-SNE to identify clusters time-series sensor data and determine four patterns, namely (1) stop and go driving, (2) changing lanes abruptly, (3) driving too fast for road condition, and (4) not properly maintaining lane.  A transfer learning with fairness optimization model has also been developed to optimize the fairness between different demographic groups and minimize the loss of performance.  Such model also provides explanations by identifying the important features that affect the fairness between groups.  Six PhD students and one postdoctoral student were trained in this project.  Four out of the six Phd students are female and one female student is an African American.  Two PhD students have public health or nursing backgrounds.  The interdisciplinary training has broadened their knowledge in health and computing.  

Last Modified: 01/16/2024
Modified by: Bhupesh   Shetty