Effects of Uncertainty and Cognitive Load on User Trust in Predictive Decision Making
Effects of Uncertainty and Cognitive Load on User Trust in Predictive Decision Making
复制标题
不确定性和认知负荷对预测决策中用户信任的影响
DOI：
--
复制
发表时间：
2017
期刊：
影响因子：
--
通讯作者：
Fang Chen
中科院分区：
文献类型：
--
作者：
Jianlong Zhou;Syed Arshad;Simon Luo;Fang Chen
不同领域数据的快速增长导致基于机器学习（ML）的智能系统在预测决策场景中的广泛应用。不幸的是，由于其复杂的工作机制，这些系统对用户来说就像一个“黑盒子”，因此严重影响了用户对人机交互的信任。这部分是由于机器学习模型中固有的紧密耦合的不确定性，这些不确定性是预测性决策建议的基础。此外，当这种分析驱动的智能系统用于现代复杂的高风险领域（如航空）时，除了信任之外，用户决策还受到更高水平的认知负荷的影响。本文研究了不确定性和认知负荷对预测决策中用户信任的影响，以期为基于机器学习的智能系统设计有效的用户界面。我们在重复析因设计实验中对42名受试者进行的用户研究发现，不确定性类型（风险和模糊性）和认知工作量水平都会影响预测决策中的用户信任。不确定性呈现导致信任增加，但只有在低认知负荷条件下，当用户有足够的认知资源来处理信息时。在高负荷条件下（当认知资源供应不足时），不确定性的出现会导致对系统及其建议的信任度下降。
Rapid increase of data in different fields has been resulting in wide applications of Machine Learning (ML) based intelligent systems in predictive decision making scenarios. Unfortunately, these systems appear like a ‘black-box’ to users due to their complex working mechanisms and therefore significantly affect the user’s trust in human-machine interactions. This is partly due to the tightly coupled uncertainty inherent in the ML models that underlie the predictive decision making recommendations. Furthermore, when such analytics-driven intelligent systems are used in modern complex high-risk domains (such as aviation) - user decisions, in addition to trust, are also influenced by higher levels of cognitive load. This paper investigates effects of uncertainty and cognitive load on user trust in predictive decision making in order to design effective user interfaces for such ML-based intelligent systems. Our user study of 42 subjects in a repeated factorial design experiment found that both uncertainty types (risk and ambiguity) and cognitive workload levels affected user trust in predictive decision making. Uncertainty presentation leads to increased trust but only under low cognitive load conditions when users had sufficient cognitive resources to process the information. Presentation of uncertainty under high load conditions (when cognitive resources were short in supply) leads to a decrease of trust in the system and its recommendations.