Feasibility Analysis of Applying Deep Neural Network on Driving Distance Estimation

In numerous location-based applications, such as the vehicular routing problem, driving distances play a crucial role. However, these driving distances often differ from the direct geographic distance computed using latitude and longitude. Hence, accurately estimating the driving distance between two positions is vital for the success of such services. Researchers have worked on developing efficient methods to estimate driving distances, and it has been reported that the inflation ratio (or detour index) between driving distance and direct distance is approximately 1.3. However, this simple method may not suffice for complex road networks. To address these challenges, other researchers have proposed deep learning based approaches. They show relatively good performance for real road data sets. Even though the deep learning based approach may work for real road data sets, it crucial to fully understand the behavior of the deep neural network (DNN) based approach. Therefore, in this study, We aim to thoroughly examine a deep learning-based road network distance estimation method under controlled conditions. Specifically, we define five different distance types and assess the performance of the DNN-based approach. Subsequently, we analyze the key factors that influence its performance. Through extensive simulations, we demonstrate that the DNN-based method performs well across most distance definitions. After conducting a thorough analysis of the evaluation results, we have identified a key characteristic of the road data sets that significantly impacts the accuracy of the DNN-based method. Specifically, we have found that the “discontinuity” of the distances plays a crucial role in achieving high accuracy. As a result, we propose that future designs of DNN-based road network distance estimation methods should prioritize careful consideration of this “discontinuity” aspect to optimize their performance and ensure better accuracy.