Computational Techniques for Studying Everyday Multiattribute Choice

日常选择对象，如食品、电影、衣服和消费品，可以被视为具有不同的属性或特征。尽管在日常物品之间的选择涉及到对这些属性的关注和评估，但这些属性本身可能是复杂的，不易被研究人员观察到。该项目试图通过将机器学习和统计学的见解与决策研究中的现有理论相结合，开发计算技术来揭示日常选择中涉及的属性。它提供了利用机器学习和数据科学的发展来推进我们对日常决策的理论理解的可能性，并在此过程中，通过为医疗保健或退休计划等重要决策的实施和政策提供信息和改进，产生更广泛的影响。这个项目有两个主要组成部分。第一个是计算性的，涉及使用统计技术从大规模用户生成的互联网数据中恢复（否则不可观察的）现实世界选择对象的属性表示。第二个主要组成部分是经验的，它涉及使用这些恢复的属性，结合现有的多属性决策规则，研究各种现实世界对象之间的多属性选择。总的来说，该项目将所提出的方法应用于三个领域：电影选择、书籍选择和食物选择，并且对于这些领域中的每一个，都试图预测涉及电影、书籍和食物的自然决策问题中的选择概率、决策时间和属性重要性的判断，这些问题都给了实验室中的参与者。以类似的方式，本项目使用这些领域来测试使用多属性研究中流行的程式化实验类型建立的选择集依赖和参考依赖等行为效应，是否也适用于考虑的对象是自然的，并且没有使用明确的属性-对象矩阵来描述。最后，本项目使用这些领域来研究选择集本身存储在内存中的决策，而不是明确地呈现给决策者。

Everyday choice objects, such as food items, movies, clothes, and consumer goods, can be seen as possessing different attributes or features. Although choices between everyday objects involve attending to and evaluating these attributes, the attributes themselves may be complex and not easily observed by researchers. This project attempts to develop computational techniques to uncover the attributes involved in everyday choice by combining insights from machine learning and statistics with existing theories in decision making research. It offers the possibility harnessing developments in machine learning and data science to advance our theoretical understanding of everyday decision making and, in the process, yield broader impacts by informing and improving implementation of and policy toward important decisions like those concerned with health care or retirement planning.There are two major components to this project. The first is computational, and involves the use of statistical techniques to recover (otherwise unobservable) attribute representations for real-world choice objects from large-scale user-generated internet data. The second major component is empirical, and involves the use of these recovered attributes, combined with existing multi-attribute decision rules, to study multi-attribute choices between various real-world objects. Overall, the project applies the proposed approach to three domains: movie choice, book choice, and food choice, and for each of these domains, attempts to predict choice probabilities, decision times, and judgments of attribute importance in naturalistic decision problems involving movies, books, and food items, given to participants in the laboratory. In a similar manner, this project uses these domains to test whether behavioral effects such as choice set dependence and reference dependence, established using the types of stylized experiments popular in multi-attribute research, also hold when the objects under consideration are naturalistic and are not described using explicit attribute-by-object matrices. Finally this project uses these domains to study decisions in which the choice sets themselves are stored in memory, and are not explicitly presented to decision makers.