Deep encoder cross network for estimated time of arrival

Abstract

Estimated time of arrival (ETA) is essential to enable various intelligent transportation services and reduce passenger waiting time. Estimating the time of arrival of public transport in a highly dynamic and uncertain transportation system could be challenging. Many indirect factors beyond the remaining travel distance could dramatically deviate the time of arrival from the original schedule. Existing distance-based estimation methods disregarding those factors usually result in inaccurate estimations. In this paper, we propose a new deep learning model, called Deep Encoder Cross Network (DECN), to improve the ETA prediction based on multiple non-distance-based factors such as weather, road speed and congestion, and traffic composition. Unlike most regression tasks that output the target directly, we predict the ETA residual over the location-based ETA prediction. To effectively learn in the large and sparse input feature space, we use a new neural network structure consisting of three main components. First, a deep neural network is responsible for modeling explicit feature interactions. Second, an encoder network is constructed to reduce the input feature dimensionality. Third, a cross-network is introduced to learn from the implicit feature interactions. We conduct extensive experiments on a large real-world bus ETA dataset of Hong Kong, which contains about 2.95 × 108 rows with 27 different features on an 84-dimensional space. The results show that the deep learning approach with the DECN model can improve the ETA error by 11% on average, and 49% for late arrival. The proposed approach can be further improved and extended to estimate other traffic information by incorporating non-distance-based related information.