Speeding up $KNN−\overline{W}H$ for Origin–Destination Travel Time Estimation

Sebastian Moreno, Sofia Alvarez, WILFREDO F YUSHIMITO, Alfonso Tobar-Arancibia

Keywords: origin–destination travel time, machine learning, k-nearest neighbor, speed up estimation process

TL;DR: We propose k-KNN-WH, a two-step framework for OD time prediction that clusters the data using 𝑘-means and then applies KNN-WH in the corresponding cluster. Empirical results show minimal impact on MAPE while reducing the time estimation process.

Abstract: Origin-destination (O-D) travel time estimation is among the most important problems studied in transportation. It focuses on determining accurate travel time from a specific origin point to a destination point. Given the development of new technologies such as GPS and mobile applications, this data can be easily gathered, improving the estimation of the O-D travel time and enabling prediction in almost real-time. Currently, one of the simplest and newest algorithms is the $KNN-\overline{W}H$ model, an improvement of the K-Nearest Neighbors method with Haversine distance and a correction factor. Unfortunately, the direct application of this method can take over 50 minutes to predict a new set of 70,000 data points. This paper proposes $k-KNN-\overline{W}H$, a new two-step framework that clusters the data using $k$-means and then applies $KNN-\overline{W}H$ on the corresponding cluster. The empirical results show a minimal impact on the MAPE performance $(1.5\%)$ while reducing the time estimation process from approximately 50 to 20 minutes.

Submission Number: 4