CAREER: Interpretable Deep Modeling of Discrete Time Event Sequences

离散时间事件序列（DTES）是有序的事件序列，每个事件都有一个具体的时间戳。DTES在我们的日常生活中无处不在。一个典型的例子是病人的电子健康记录。DTES的计算建模可以揭示隐藏事件的演化机制，提高序列预测和分组等端点分析任务的性能。用于分析DTES的传统方法通常基于强统计假设，并且在实践中可能不起作用。受最近深度学习方法在各个应用领域的经验成功的启发，该项目的目标是开发可解释的深度学习方法来建模DTES。该项目验证了所开发的算法在各种医疗应用中的实用性。它将由此产生的研究成果纳入课程开发和课程，以培养新一代机器学习和数据挖掘从业人员。此外，还为高中生和社区大学生提供特别培训机会，以便更广泛地教育现代数据分析技术。首先，它开发了一系列的方法集成外部领域的知识到建模过程中。这保证了学习的模型与领域知识很好地对齐，同时提供了有效的正则化以避免过拟合。其次，设计了基于模仿学习和模式分解的方法来解释隐藏在学习模型中的知识。这使得学习的模型更加实用和可重用。第三，开发有效的模型和数据共享机制，以在类似的学习任务之间转移知识。这通过利用任务关系来最大化每个任务的可用样本的利用率。在医学领域的两个关键问题，医院读取使命和疾病表型，被用来作为目标应用程序验证所提出的方法的基础上，几个现实世界的大规模患者电子健康记录datasets.This奖项反映了NSF的法定使命，并已被认为是值得通过使用基金会的智力价值和更广泛的影响审查标准进行评估的支持。

Discrete Time Event Sequences (DTES) are ordered event sequences with a concrete timestamp associated with each event. DTES are ubiquitous in our daily life. One representative example is patient electronic health records. Computational modeling of DTES can reveal the hidden event evolving mechanisms and improve the performance of endpoint analytical tasks such as sequence forecasting and grouping. Conventional approaches for analyzing DTES are typically based on strong statistical assumptions and may not work well in practice. Motivated by the recent empirical success of deep learning methods in various application domains, the objective of this project is to develop interpretable deep learning approaches for modeling DTES. This project validates the utility of the developed algorithms in various medical applications. It incorporates the resulting research outcomes into curriculum development and courses, to train a new generation of machine learning and data mining practitioners. In addition, special training opportunities are provided to high school students and community college students for a broader education of modern data analysis techniques.This project consists of three synergistic research thrusts. First, it develops a series of approaches for integrating external domain knowledge into the modeling process. This guarantees the learned models align well with the domain knowledge and at the same time provides effective regularizations to avoid overfitting. Second, it devises approaches based on mimic learning and pattern dissection to interpret the knowledge hidden in the learned models. This makes the learned models much more practical and reusable. Third, effective model and data sharing mechanisms are developed to transfer the knowledge across similar learning tasks. This maximizes the utilizations of the available samples for each task by leveraging the task relationships. Two key problems in medical domain, hospital readmission and disease phenotyping, are used as the target applications for validating the proposed approaches based on several real-world large-scale patient electronic health record data sets.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.