Applying Machine Learning Algorithms to Highway Safety EEPDO

摘要

Estimating expected equivalent property damage only (EEPDO) is crucial to highway crash hotspot identification (HSID), which is a key component of a highway safety improvement program. During the past 60 years, HSID methodologies advanced steadily from traditional scan based methods to statistical model based methods and have reached the sophistication of encompassing advanced statistical models with many variations and refinements while there still exist a number of theoretical issues unsolved. Consequently, these advanced models are not widely used in the practice of transportation engineering. This paper investigated the performance of an easy to use alternative to estimate the EEPDO -- using machine learning techniques of K nearest neighbor (KNN) algorithm and compared it against the prevalent statistical model -- Negative Binomial (NB). NB assumes that the raw data follow a certain Gamma distribution which is not ubiquitously hold for crash data. Comparatively, being a nonparametric predictor, KNN is expected to produce better estimation on crash data in that it requires no assumption on the raw data. For experiment, a case study was conducted on highway US 49 in Harrison County of Mississippi. The results indicated that KNN outperformed NB.