Wait Time Prediction For Airport Taxis Using Weighted Nearest Neighbor Regression

In this paper, we address the neighborhood identification problem in the presence of a large number of heterogeneous contextual features. We formulate our research as a problem of queue wait time prediction for taxi drivers at airports and investigate heterogeneous factors related to time, weather, flight arrivals, and taxi trips. The neighborhood-based methods have been applied to this type of problem previously. However, the failure to capture the relevant heterogeneous contextual factors and their weights during the calculation of neighborhoods can make existing methods ineffective. Specifically, a driver intelligence-biased weighting scheme is introduced to estimate the importance of each contextual factor that utilizes taxi drivers' intelligent moves. We argue that the quality of the identified neighborhood is significantly improved by considering the relevant heterogeneous contextual factors, thus boosting the prediction performance (i.e., mean prediction error < 0.09 and median prediction error < 0.06). To support our claim, we generated an airport taxi wait time dataset for the John F. Kennedy International Airport by fusing three real-world contextual datasets, including taxi trip logs, passenger wait times, and weather conditions. Our experimental results demonstrate that the presence of heterogeneous contextual features and the drivers' intelligence-biased weighting scheme significantly outperform the baseline approaches for predicting taxi driver queue wait times.