Machine Learning Tools for Informing Transportation Technology and Policy

US Transportation Collection2019PDFTech ReportCummings, MaryLi, SongpoDuke UniversityCollaborative Sciences Center for Road SafetyUnited States. Department of Transportation. Federal Highway AdministrationCollaborative Sciences Center for Road SafetyUnited StatesAlternatives analysisData analysisEvaluation and assessmentFatalitiesLogistic regression analysisMachine learningNeural networksExplainable artificial intelligenceFeature selectionCSCRS-R10Collaborative Sciences Center for Road Safety (Grant #: 69A3551747113)Final Report (March 2018-November 2019)With rapid advances in data analytic tools, machine learning techniques are increasingly being used to study large transportation crash data sets. The objective of such approaches is to reveal underlying and potentially unknown patterns of influence between driver and pedestrian characteristics, environment factors, vehicle attributes and crash fatalities. However, machine learning results can be greatly affected by the subjectivity of the machine learning practitioner, where the practitioner subjectively selects the machine learning algorithm and algorithm parameters for a specific data set, and either this person or perhaps other people then interpret the results. Little work has been conducted to study this practitioner-induced subjectivity problem in order to understand its causes, influences, and methods for avoidance, particularly in transportation settings. To help fill this gap, two transportation datasets examining driver and pedestrian accident fatalities were analyzed with two different machine learning techniques of low and high complexity (logistic regression and neural networks). The results demonstrate that both the type of model and feature interpretation method produce different results in terms of model performance and assessment of feature importance. These outcomes that highlight more than ten opportunities for subjective decisions suggest that more work is needed in looking at how such subjective modeling and interpretation choices affect the use of machine learning models in support of policy decision making.113