A stepwise interpretable machine learning framework using linear regression (LR) and long short-term memory (LSTM): City-wide demand-side prediction of yellow taxi and for-hire vehicle (FHV) service
A stepwise interpretable machine learning framework using linear regression (LR) and long short-term memory (LSTM): City-wide demand-side prediction of yellow taxi and for-hire vehicle (FHV) service
复制标题
DOI：
10.1016/j.trc.2020.102786
复制
发表时间：
2020-11-01
影响因子：
8.3
通讯作者：
Pendyala, Ram M.
中科院分区：
文献类型：
--
作者：
Kim, Taehooie;Sharda, Shivam;Pendyala, Ram M.
随着基于应用程序的叫车服务在现有的传统出租车市场中被广泛采用，研究人员一直致力于了解影响新移动需求的重要因素。计量经济模型（EM）主要用于解释需求的重要因素，深度神经网络（DNN）最近已被用于通过捕获大型数据集中的复杂模式来提高预测性能。然而，为了缓解可能的（诱发的）交通拥堵，平衡当前出租车司机的使用率，仍然迫切需要一种积极主动地管理新兴服务和常规出租车配额制度的有效战略。本文旨在系统地设计一个可解释的深度学习模型，该模型能够评估平衡两种模式之间需求量的配额系统。通过线性回归（LR）模型，结合长短期记忆（LSTM）分层的神经网络，开发了一个两阶段可解释的机器学习建模框架。第一阶段调查现有出租车和按需叫车服务之间的相关性，同时控制其他解释变量。第二阶段实现长短期记忆（LSTM）网络结构，捕获来自第一估计阶段的残差，以提高预测性能。提出的逐步建模方法（LR-LSTM）预测出租车乘车需求，并使用纽约市（NYC）的出租车数据在接送需求预测的应用程序中实现。实验结果表明，该集成模型可以捕捉现有出租车和叫车服务之间的相互关系，并识别其他因素的影响，即星期几，天气和假期。总的来说，这种建模方法可以应用于构建一个有效的主动需求管理（ADM）的短期内，以及按需叫车服务和传统出租车之间的配额控制策略。
As app-based ride-hailing services have been widely adopted within existing traditional taxi markets, researchers have been devoted to understand the important factors that influence the demand of the new mobility. Econometric models (EMs) are mainly utilized to interpret the significant factors of the demand, and deep neural networks (DNNs) have been recently used to improve the forecasting performance by capturing complex patterns in the large datasets. However, to mitigate possible (induced) traffic congestion and balance utilization rates for the current taxi drivers, an effective strategy of proactively managing a quota system for both emerging services and regular taxis is still critically needed. This paper aims to systematically design an explainable deep learning model capable of assessing the quota system balancing the demand volumes between two modes. A two-stage interpretable machine learning modeling framework was developed by a linear regression (LR) model, coupled with a neural network layered by long short-term memory (LSTM). The first stage investigates the correlation between the existing taxis and on-demand ride-hailing services while controlling for other explanatory variables. The sec ond stage fulfills the long short-term memory (LSTM) network structure, capturing the residuals from the first estimation stage in order to enhance the forecasting performance. The proposed stepwise modeling approach (LR-LSTM) forecasts the demand of taxi rides, and it is implemented in the application of pick-up demand prediction using New York City (NYC) taxi data. The experiment result indicates that the integrated model can capture the inter-relationships between existing taxis and ride-hailing services as well as identify the influence of additional factors, namely, the day of the week, weather, and holidays. Overall, this modeling approach can be applied to construct an effective active demand management (ADM) for the short-term period as well as a quota control strategy between on-demand ride-hailing services and traditional taxis.