Investigating the Role of Dimension Reduction in Counteracting Machine Learning Performance Bias in TSMO Strategies

Machine learning models are increasingly being utilized to develop predictive models for Transportation Systems Management and Operations (TSMO) applications. These models are often assessed based on a global performance metric that evaluates the model's performance when the entire testing dataset is presented to the model. A TSMO application is expected to perform reliably and consistently in various situations and roadway conditions. The reliability and consistency of the model predictions for various scenarios are critical to the success of transportation agencies' efforts to address mobility and safety issues. Performance bias might be influencing the model when a model's performance is inconsistent for different scenarios. This paper investigates the performance bias that the traffic management center may face when applying machine learning methods to predict incident clearance time. Additionally, dimension-reduction techniques are employed as mitigation techniques in the model development process. This paper investigates the impact of two common dimension-reduction methods, important feature selection and principal component analysis, on performance bias. In a case study, this paper investigates the performance bias of RF, BRNN, SVR, KNN, XGB, GP, and NNET in incident clearance time prediction. Incident data from three interstate corridors in Missouri, USA, were utilized to develop and evaluate the models. Repeated k-fold cross-validation was used to prepare 20 training and testing sets to demonstrate and assess the learners' performance variations due to data splits. The results indicated that the seven learners suffered from performance bias. The analysis of the impact of dimension-reduction models revealed that the important feature selection method did not significantly mitigate the performance bias. On the other hand, the principal component analysis method significantly mitigated this bias for all learners, with poor-performing learners gaining the most improvements. In addition to contributing to reducing the performance bias, the principal component analysis significantly reduced the learners' global (i.e., overall) error metrics.