Co-MEAL: Cost-Optimal Multi-Expert Active Learning Architecture for Mobile Health Monitoring
Co-MEAL: Cost-Optimal Multi-Expert Active Learning Architecture for Mobile Health Monitoring
复制标题
Co-MEAL：用于移动健康监测的成本最优的多专家主动学习架构
DOI：
10.1145/3107411.3107430
复制
发表时间：
2017
期刊：
影响因子：
--
通讯作者：
A. Gebremedhin
中科院分区：
文献类型：
--
作者：
Ramyar Saeedi;K. Sasani;A. Gebremedhin
移动的健康监测在各种医疗保健应用中发挥着核心作用。使用移动的技术，医疗保健提供者可以实时访问临床信息并与受试者进行交流。由于医疗保健应用的敏感性，这些系统需要高度准确地处理生理信号。然而，由于在动态环境中采用移动的设备，因此每当系统的配置发生变化时，机器学习模型的准确性就会下降。因此，需要专门解决与动态环境（例如，不同用户、信号异质性）相关联的挑战的数据挖掘和机器学习技术。在本文中，使用主动学习作为组织原则，我们提出了一个成本最优的多专家架构，以适应在给定的上下文中开发的机器学习模型（例如分类器），以一个新的上下文或配置。更具体地说，在我们的架构中，系统的机器学习模型从系统可用的专家（例如另一个移动终端，人类注释者）学习，同时最大限度地减少数据标记的成本。我们的架构还利用专家之间的协作来丰富他们的知识，这反过来又降低了未来步骤中数据标记的成本和不确定性。我们使用公开的人类活动数据集证明了该架构的有效性。我们表明，活动识别的准确率达到85%以上，标记只有15%的未标记的数据。与此同时，来自人类专家的查询数量减少了82%。
Mobile health monitoring plays a central role in a variety of health-care applications. Using mobile technology, health-care providers can access clinical information and communicate with subjects in real-time. Due to the sensitive nature of health-care applications, these systems need to process physiological signals highly accurately. However, as mobile devices are employed in dynamic environments, the accuracy of a machine learning model drops whenever a change in configuration of the system occurs. Therefore, data mining and machine learning techniques that specifically address challenges associated with dynamic environments (e.g. different users, signal heterogeneity) are needed. In this paper, using active learning as an organizing principle, we propose a cost-optimal multiple-expert architecture to adapt a machine learning model (e.g. classifier) developed in a given context to a new context or configuration. More specifically, in our architecture, a system's machine learning model learns from experts available to the system (e.g. another mobile device, human annotator) while minimizing the cost of data labeling. Our architecture also exploits collaboration between experts to enrich their knowledge which in turn decreases both cost and uncertainty of data labeling in future steps. We demonstrate the efficacy of the architecture using a publicly available dataset on human activity. We show that the accuracy of activity recognition reaches over 85% by labeling only 15% of unlabeled data. At the same time, the number of queries from human expert is reduced by up to 82%.