Identifying Nearest Neighbors in a Large-Scale Incident Data Archive

Large-scale data archives compiled by transportation management systems capture a wide range of experience. However, to use this experience effectively, transportation professionals must be able to search these archives efficiently to identify cases of interest. An example of this challenge is to find past events that are similar to a current event to understand traffic flow and the impact of utilized control strategies. These similar events are referred to as nearest neighbors to the current event. To identify "near" neighbors in a data archive, a distance metric is required to measure similarity between cases. However, common distance metrics, such as Euclidean distance, are not valid when used with categorical variables, which commonly make up traffic event data. To address this challenge, this research effort developed a distance metric that can be effectively used with categorical data. The metric is based on the influence of variable values on a measurable objective meaningful to the purpose of selecting nearest neighbors. The metric was developed in a case study to identify similar past incidents in an archive to a current incident, with the objective of forecasting incident duration. When this method was incorporated in a nonparametric regression forecasting model, it was demonstrated to outperform parametric forecasting models significantly.