标题： 将离散选择模型构造成具有效用解释的深度神经网络 显示英文标题

标题： Framing Discrete Choice Model as Deep Neural Network with Utility Interpretation

摘要： 深度神经网络（DNN）已被越来越多地应用于交通需求预测。然而，尚无研究探讨DNN与基于效用的离散选择模型（DCM）之间的关系，而不仅仅是对预测准确性进行简单比较。为了填补这一空白，本文从理论角度探讨了DNN与DCM之间的关系，并得出了三个主要发现。首先，我们将效用解释引入到DNN模型中，并证明DCM是具有浅层和稀疏结构、可识别参数、逻辑损失、零正则化和基于领域知识的特征转换的DNN的一个特例。其次，一系列四个神经网络模型展示了在交通方式选择背景下，DNN如何逐步放弃可解释性以换取预测能力。通过强大的表示学习和高模型容量，DNN实现了高预测能力；但通过失去凸优化和统计特性以及参数的不可识别性，牺牲了可解释性。第三，效用解释使我们能够开发一种数值方法，从DNN中提取重要的经济信息，包括选择概率、弹性、替代边际率和消费者剩余。总体而言，本研究做出了三个贡献：理论上，它将DCM框架作为DNN的一个特例，并将效用解释引入DNN；方法上，它展示了DCM与DNN之间的可解释性-预测性权衡，并提出了它们联合改进的潜力；实践中，它引入了一种事后数值方法，从DNN中提取经济信息，并通过效用概念使其可解释。 显示英文摘要

摘要： Deep neural network (DNN) has been increasingly applied to travel demand prediction. However, no study has examined how DNN relates to utility-based discrete choice models (DCM) beyond simple comparison of prediction accuracy. To fill this gap, this paper investigates the relationship between DNN and DCM from a theoretical perspective with three major findings. First, we introduce the utility interpretation to the DNN models and demonstrate that DCM is one special case of DNN with shallow and sparse architecture, identifiable parameters, logistic loss, zero regularization, and domain-knowledge based feature transformation. Second, a sequence of four neural network models illustrate how DNN gradually trade away interpretability for predictability in the context of travel mode choice. High predictability is achieved by DNN's powerful representation learning and high model capacity; but interpretability is sacrificed through the loss of convex optimization and statistical properties, and non-identification of parameters. Third, the utility interpretation allows us to develop a numerical method of extracting important economic information from DNN including choice probability, elasticity, marginal rate of substitution, and consumer surplus. Overall, this study makes three contributions: theoretically it frames DCM as a special case of DNN and introduces the utility interpretation to DNN; methodologically it demonstrates the interpretability-predictability tradeoff between DCM and DNN and suggests the potential of their joint improvement, and practically it introduces a post-hoc numerical method to extract economic information from DNN and make it interpretable through the utility concept.