Leveraging Machine Learning in Modern Revenue Management

近十年来，收入管理在获取更丰富的个人数据和实施复杂算法方面取得了巨大增长。例如，基于社交网络的个性化推荐和推荐程序已经成为在线零售商的常见做法。然而，黑盒机器学习算法在帮助公司进行指令性决策分析、为经理生成可解释的报告以及让客户和监管机构相信没有制定歧视性政策方面做得不够。在这项提议中，我们计划利用尖端机器学习算法的卓越实用性能，如随机森林、梯度提升和高斯过程，并提供可解释的建模框架，以连接机器学习和收入管理的最佳部分。首先，我们提出了一种新的客户选择模型--二叉决策树，并将树的聚集与随机森林联系起来。该模型可以让零售商解码各种类型客户的购买模式，如搜索和替代效应，并在面对新设计的产品时准确预测客户行为。其次，我们提出用高斯过程来建模客户之间的社会交互。在机器学习和贝叶斯优化中，通常使用高斯过程来调整超参数。它们允许融合复杂和细微差别的社会关系以及不同的客户类型。作为一种易于管理的有监督学习算法，GP回归框架可以更深入地理解客户的社会活动是如何诱导购买的，并最终提高企业营销策略和收入的有效性。第三，我们建议将梯度增强回归树应用于客户对个性化产品推荐的反应的建模和预测。梯度提升已被证明在算法交易和生存分析中极为成功。在个性化产品推荐中，回归树将提供客户和产品功能交互的简单而强大的视图，从而转换为高性能推荐策略。通过这项拟议的研究，我们希望利用机器学习算法，这些算法最初是为“预测性”任务设计的，并将它们转变为强大的主力，为公司和监管机构提供“指令性”政策建议。从公司的角度来看，我们希望在我们的框架下，准确预测的百分比-评估机器学习算法的常见性能指标-可以转化为收入和利润的显著增长。从监管机构的角度来看，我们希望拟议的研究将在复杂的算法和公平透明的商业做法之间提供一座桥梁。

In the recent decade, revenue management has seen tremendous growth in the acquisition of richer personal data and the implementation of sophisticated algorithms. For example, personalized recommendations and referral programs based on social networks have become common practices of online retailers. However, black-box machine learning algorithms fall short in helping firms conducting prescriptive decision analytics, generating interpretable reports for managers, and convincing customers and regulators that no discriminatory policies are devised. In this proposal, we plan to leverage the outstanding practical performance of cutting-edge machine learning algorithms such as random forests, gradient boosting and gaussian processes and provide interpretable modeling frameworks to bridge the best parts of machine learning and revenue management. First, we propose a novel model of customer choices as binary decision trees, and connect the aggregation of trees to random forests. The model may allow retailers to decode the purchasing patterns of various types of customers such as searching and substitution effects, and accurately predict customer behavior when facing newly designed products. Second, we propose to model social interaction between customers by Gaussian processes. Gaussian processes are commonly used in machine learning and Bayesian optimization to tune hyper-parameters. They allow to incorporate complex and nuanced social ties as well as heterogeneous customer types. As a tractable supervised learning algorithm, the GP-regression framework may lead to deeper understanding of how customers' social activities induce purchasing and eventually increase the efficacy of firms' marketing strategies and revenues. Third, we propose to apply gradient boosting regression trees to the modeling and prediction of customers' reactions to personalized product recommendations. Gradient boosting has been shown to be extremely successful in algorithmic trading and survival analysis. In personalized product recommendations, the regression trees will provide a simple and powerful view of the interactions of customer and product features and thus be translated to high-performance recommendation policies. By the proposed study, we hope to leverage machine learning algorithms, which are originally designed for "predictive" tasks, and transform them to powerful workhorses for "prescriptive" policy recommendations for firms and regulators. From the firms' point of view, we hope that under our framework, the percentage of accurate predictions, a common performance indicator in evaluating machine learning algorithms, can be translated to the significant increase in revenues and profits. From the regulators' point of view, we hope that the proposed study will provide a bridge between sophisticated algorithms and fair and transparent business practices.