CHS: Medium: Behavior360: Learning a Human Behaviorome in Uncontrolled Settings

随着移动传感、计算和机器学习的成熟，科学家们现在可以构建智能工具，通过分析来自商业、可穿戴传感器（如智能手表）的数据来更好地理解人类行为。该项目的目标是设计、构建和评估新的算法，这些算法可以从智能手表传感器在野外连续收集的数据中持续感知、建模、分析和解释人类行为。机器学习方法将被设计用于从收集的数据中推断行为模式，并生成行为模式的解释，这些解释可以被具有不同背景的人轻松解释。计算方法将使用历史数据以及在自由生活环境中收集的新数据进行评估。健康和行为之间的关系将通过使用机器学习从收集和建模的传感器数据中获得临床健康评分来探索。这项研究将涉及来自不同学科和人口背景的学生，通过参与暑期研究项目和顶点项目。本项目的技术目标分为三个目标。首先，由于持续的行为感知需要的资源超过了当前智能手表的功率容量，研究人员将设计算法，优化预测性能和功耗之间的权衡。其次，研究人员将创建鲁棒的行为模型，将稀疏标签信息与传感器数据结合起来，自动构建人类行为词汇表，并采用领域自适应来泛化跨越人物、时间和行为类型的模型。第三，研究人员将创建机器学习方法，通过自动生成的文本解释，从行为数据中生成准确且可解释的临床健康评分。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

With the maturing of mobile sensing, computing, and machine learning, scientists can now build intelligent tools to better understand human behavior by analyzing data from commercial, wearable sensors (such as smartwatches). The goal of this project is to design, build, and evaluate novel algorithms that continuously sense, model, analyze, and interpret human behavior from smartwatch sensor data collected continuously in the wild. Machine learning methods will be designed to infer behavior patterns from collected data as well as generate explanations of behavior patterns that can be easily interpreted by humans with diverse backgrounds. The computational methods will be evaluated using historical data as well as new data collected in free-living environments. The relationship between health and behavior will be explored by using machine learning to derive clinical health scores from collected and modeled sensor data. The research will involve students from diverse disciplinary and demographic backgrounds through involvement in summer research programs and capstone projects.The technical goals of this project are divided into three aims. First, because continuous behavior sensing requires resources that exceed the power capacity of current smartwatches, the investigators will design algorithms that optimize the trade-off between predictive performance and power consumption. Second, the investigators will create robust behavior models that combine sparse label information with sensor data to automatically construct a vocabulary of human behavior, and employ domain adaptation to generalize models across people, times, and behavior types. Third, the investigators will create machine learning methods to produce accurate and interpretable clinical health scores from behavior data with automatically-generated text explanations.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.