Taxi is one of the most important ways for citizens' daily travel, but taxi service faces a typical problem that greedy drivers may deliberately take unnecessary detours to overcharge passengers. An in-depth analysis of drivers' detour behavior is necessary to ensure high-quality service. In this paper, two kinds of detour patterns, namely kind detours and malicious detours, are defined and identified based on taxi datasets collected from three metropolitan cities in two countries. To better understand the detour choices of drivers, we explore the factors that may influence different detour patterns in terms of drivers, spatio-temporal distribution, land use, and network characteristics, and find that these two types of detours have distinctly different features. Based on these analyses, the detour behavior is modeled as a multi-class problem taking into account various features such as actual time, driver trip grids, driver average daily trips, origin/destination trip degrees, origin/destination land use, etc. Considering that our dataset is imbalanced due to significantly fewer detour trips than normal driving trips, a combined model of hybrid sampling and ensemble learning is used to predict detour choices at the beginning of the trip. Results show that the proposed method is useful and powerful in the prediction of detour behavior. This paper is a quantitative study to empirically reveal the factors influencing different detour patterns and to perform ex ante predictions of detour choices, which facilitates managers to understand detour behavior and develop appropriate interventions.

出租车是市民日常出行最重要的出行方式之一，但出租车服务面临的一个典型问题是，贪心的司机可能故意走不必要的弯路，向乘客多收费用。对司机的绕行行为进行深入分析是确保优质服务的必要条件。在本文中，基于从两个国家的三个大城市收集的出租车数据集，定义和识别了两种绕行模式，即善意绕行和恶意绕行。为了更好地理解驾驶员的绕行选择，我们从驾驶员、时空分布、土地利用和网络特征等方面探讨了可能影响不同绕行模式的因素，发现这两类绕行具有明显不同的特征。基于这些分析，绕行行为被建模为一个多类问题，考虑到各种特征，如实际时间、司机出行网格、司机平均每日出行、起点/目的地出行度、起点/目的地土地使用等。考虑到我们的数据集是不平衡的由于绕行次数明显少于正常驾驶行程，因此使用混合采样和集成学习的组合模型来预测行程开始时的绕行选择。结果表明，所提出的方法在预测绕行行为方面是有用且强大的。本文是一项定量研究，旨在通过实证揭示影响不同绕行模式的因素，并对绕行选择进行事前预测，从而有助于管理者了解绕行行为并制定适当的干预措施。