Collaborative Research: CISE-MSI: DP: III: Information Integration and Association Pattern Discovery in Precision Phenomics

该项目解决了一类涉及松耦合异构域的机器学习问题。这些问题在各个领域都有重要的应用，从精准农业（植物基因型到表型）到精准医学（基因型-疾病关联）到生态学（土壤水分-气候相互作用）。除了人工智能（AI）和机器学习方面的具体挑战之外，一个同样重要且相关的挑战是使人才库多样化，并具备解决这些问题的技术能力。该项目有三个具体目标：(1)开发在松耦合域中发现关联模式的方法；(2)对开发的方法进行信息集成、原型开发和评价；(3)西弗吉尼亚州立大学（WVSU）机器学习/人工智能及其在精准农业中的应用的研究和教育能力建设。该项目包括两个案例研究。首先是精准农业中的基因型-表现型问题，特别是重要水果作物哈瓦那辣椒的果实表现型问题。二是将提出的学习模型应用于将水果表型映射到感官知觉的问题，这是预测水果作物的味道和风味所需要的。该项目还包括能力建设工作，以提高西弗吉尼亚州立大学教师的研究能力，同时为学生提供研究培训。教育活动包括为本科生和研究生开发数据素养课程，以及机器学习和数据分析的研究培训。此外，该项目将教育当地地区的教师、学生和公众，让他们了解人工智能（AI）及其应用的创新和趋势。该项目包括精心设计的计划，包括来自西弗吉尼亚州立大学的少数民族学生和教师，这是一所HBCU（历史上的黑人学院和大学），以及西弗吉尼亚大学（WVU）的合作者，这是一所R1机构。通过这样做，该项目将提高整个WVSU大学的数据科学素养，并培养更多学生在数据驱动分析方面的技能。在看似不同的领域之间识别可靠的关联模式仍然是数据科学和机器学习基础上的核心挑战。当涉及的领域非常复杂、异构并且拥有庞大的数据集时，这样做是一项更加艰巨的任务。该项目确定了涉及松耦合异构领域的一类问题，并提出了一个创新的框架，用于跨此类复杂领域执行大规模关联发现。我们的框架建立在一个强大的理论基础之上，即根植于信息论的信息瓶颈。该项目通过开发跨两个异构领域学习隐藏关联的新方法，以及跨这些领域集成信息的新方法，为数据科学和机器学习的基础做出了新的理论贡献。该项目还开发了新的学习方法在精准农业和作物表型学两个具体问题上的新应用。所提出的在松散耦合域的已识别类下发现关联模式的方法可以很容易地针对该已识别类的任何成员进行修改。这门课相当广泛，包括人类努力的各个领域的问题，从农业到精准医学再到生态学。本项目由CISE MSI研究扩展计划和促进竞争性研究的既定计划（EPSCoR）共同资助。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。

This project addresses a general class of machine learning problems involving loosely-coupled heterogeneous domains. Such problems have significant applications in various fields, from precision agriculture (plant genotype-to-phenotype) to precision medicine (genotype-disease associations) to ecology (soil moisture-climate interactions). Beyond the specific challenges in artificial intelligence (AI) and machine learning, an equally important and related challenge is diversifying the talent pool with the technical capability to these problems. The project has three specific goals: (1) develop methods for association pattern discovery in loosely-coupled domains; (2) perform information integration, prototype development, and evaluation of the methods developed; (3) capacity building in research and education in machine learning/artificial intelligence and their applications in precision agriculture at West Virginia State University (WVSU). The project includes two case studies. The first is on the genotype-to-phenotype problem in precision agriculture, specifically on fruit phenotypes in habanero pepper, an important fruit crop. The second will apply proposed learning models to the problem of mapping fruit phenotype to sensory perception, which is needed to predict taste and flavor in fruit crops. The project also involves capacity building efforts to improve research capability for faculty at WVSU, while providing research training for students. Educational activities include curriculum development in data literacy and research training in machine learning and data analytics for undergraduate and graduate students. Further, the project will educate faculty, students, and the public within the local region on innovations and trends in artificial intelligence (AI) and its applications. The project includes deliberate plans to involve minority students and faculty from WVSU, an HBCU (Historically Black College and University) with collaborators at West Virginia University (WVU), an R1 institution. By so doing, the project will improve data science literacy across the entire university at WVSU and to develop more of the students’ skills in data-driven analysis. Identifying reliable patterns of association between seemingly disparate domains remains a core challenge at the foundation of data science and machine learning. Doing this when the domains involved are each complex, heterogeneous, and have enormous datasets is an even more arduous task. The project identifies a class of problems involving loosely-coupled heterogeneous domains, and propose an innovative framework for performing large-scale association discoveries across such complex domains. Our framework is built on a strong theoretical foundation, namely, the information bottleneck rooted in information theory. The project makes new theoretical contributions to the foundation of data science and machine learning by developing new methods for learning hidden associations across two heterogeneous domains, and for integrating information across such domains. The project also develops novel applications of the new learning methods to two specific problems in precision agriculture and crop phenomics. The proposed method of discovering association patterns under the identified class of loosely-couple domains can easily be modified for any member of this identified class. This class is quite general, and includes problems in various fields of human endeavor, from agriculture to precision medicine to ecology.This project is jointly funded by the CISE MSI Research Expansion Program and the Established Program to Stimulate Competitive Research (EPSCoR).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.