EAGER: A Graphical Approach for Choice Modeling

运营经理感兴趣的一个中心问题是，在提供特定种类的产品时，如何使用历史销售数据来准确预测收入。这样的预测随后被用于做出重要的商业决策，如分类计划、新产品开发和需求估计。选择模型被广泛用于对潜在的客户行为进行建模。传统的选择模型要么过于简单，无法准确反映人们如何做出决策的本质，要么过于复杂，以至于要么在计算上很难将模型与历史数据相适应，要么随后使用它来做出商业决策。这个早期概念探索性研究奖助金(EARGER)项目研究了创新和新颖的选择模型，这些模型被设计为在计算上高效地解决收入管理中感兴趣的决策问题，同时具有强大的预测能力。研究结果将有助于在降低运营成本的同时做出更好的商业决策。它将在重要的商业应用中提供关键技术，如品种规划、新产品开发、品牌价值评估、需求估计、最优定价和收入管理。它将在运营管理、经济学、认知心理学和机器学习等学科之间产生协作。研究部分与教育计划紧密结合，包括一门关于概率图形模型的研究生课程。本科生和研究生都将通过从事研究和应用知识来解决现实世界的问题，从而从研究活动中受益。传统选择模型的次优权衡是因为这些模型在设计时没有考虑计算效率。在这个数据规模急剧增长的时代，计算效率是首要关注的问题。该项目将建立在基于图表的概率模型的基础上，如图表上的随机行走和概率图表模型，并将导致(A)开发新的基于图表的模型，用于为计算效率而设计的选择建模；(B)开发从历史购买数据中学习这些模型的新方法；(C)开发用于根据这些模型预测客户偏好的新推理算法；以及(D)开发利用这些模型解决收入管理中的优化问题的新方法。这项研究将为一种新的基于图的收入管理建模方法奠定基础。这项工作的意义和新颖性在于，选择模型的设计目标与经济学家和认知心理学家使用的传统标准(如描述潜在理性决策过程的函数形式)有很大不同，后者没有考虑收益管理中解决决策问题的计算效率。与此相反，基于海量现代数据集进行决策的选择模型应该在设计时将计算效率嵌入到模型中。该项目开发的理论和模型将汇集概率论和图论的思想和技术，以共同推理不确定性和复杂性(如概率图形模型和图上的随机游动)，以及来自收入管理最新进展的见解和工具(如使用马尔科夫链进行选择建模)。这项研究有可能促进我们对人们在面临许多选择时如何做出决定的基本理解。

A central problem of interest to operations managers is how to use historical sales data to accurately predict revenues when offering a particular assortment of products. Such predictions are subsequently used in making important business decisions such as assortment planning, new product development, and demand estimation. Choice models are widely used in modeling the underlying customer behavior. Traditional choice models are either too simple to accurately reflect the nature of how people make decisions, or so complex that it is either computationally intractable to fit the model to historical data or to subsequently use it to make business decisions. This EArly-concept Grant for Exploratory Research (EAGER) project studies innovative and novel choice models that are designed to be computationally efficient for the decision problems of interest in revenue management and at the same time have strong predictive power. The results of the research will enable improved business decisions to be made while simultaneously reducing operational costs. It will provide key technologies in important business applications of assortment planning, new product development, brand value evaluation, demand estimation, optimal pricing, and revenue management. It will generate collaborations among the disciplines of operations management, economics, cognitive psychology, and machine learning. The research component is tightly integrated with the education plan, including a graduate course on probabilistic graphical models. Both undergraduate and graduate students will benefit from the research activities by engaging in research and applying the knowledge to solve real world problems.The suboptimal tradeoff of traditional choice models is due to the fact that these models are designed without computational efficiency in mind. In this era of tremendous increase in the scale of data being generated, computational efficiency is of primary concern. This project will build on graph-based probabilistic models such as random walks on graphs and probabilistic graphical models, and will lead to (a) development of new graph-based models for choice modeling designed for computational efficiency; (b) development of new methodologies for learning these models from historical purchase data; (c) development of novel inference algorithms for predicting the customer preferences from these models; and (d) development of new methodologies for solving optimization problems in revenue management with these models. The research will lay foundations of a new graph-based modeling approach for revenue management. The significance and novelty of the work lie in the fact that the design objective of the choice modeling is critically different from the traditional criteria used by economists and cognitive psychologists (such as describing the functional form of the underlying rational decision processes), which does not consider the computational efficiency of solving decision problems in revenue management. In contrast to this, choice models for making decisions based on massive modern datasets should have computational efficiency embedded into the models by design. The theory and models developed in this project will bring together ideas and techniques from probability theory and graph theory to jointly reason about uncertainty and complexity (such as probabilistic graphical models and random walks on graphs) as well as insights and tools from recent advances in revenue management (such as choice modeling using Markov chains). The research has a potential to advance our fundamental understanding in how people make decisions when presented with many options.