RI: Small: Multi-View Latent Class Discovery and Prediction with a Streamlined Analytics Platform

发现样本中的潜在亚群是许多科学学科中的一个重要问题。社会科学家根据行为模式识别人群中的亚群，以检验社会地位的不同影响。工程师根据性能测量来识别制造系统的故障，以检测设计缺陷。医生根据临床症状定义疾病的亚型，以确定相关的遗传风险因素。这类问题涉及两组变量：描述问题的一组描述符(例如，行为模式或症状)和一组调节因素或预测因素(例如，社会地位或遗传因素)。在没有观察到的描述符的情况下(例如，在症状出现之前)，从预测器(例如，遗传风险)准确地预测潜在类别(例如，疾病亚型)的能力将推动许多这些学科的发展。本项目旨在开发一种高效的机器学习算法平台来解决这一问题。该团队将有效地将研究和教学结合起来，让学生参与到拟议的研究中来。经过验证的方法和软件将通过项目网络知识库和科学演示得到广泛传播。该项目通过衍生新的和有效的方法来解决潜在类发现和预测问题，包括多视点协同聚类、多视点子空间聚类、多目标优化协同训练和多模式深度学习方法。将开发并行和分布式算法来实现和扩大这些方法。将构建一个简化的分析平台，以最大限度地发挥拟议方法在现实世界应用中的效用。提出的解决方案将在大规模感觉和行为数据的分析中进行评估。通过与领域专家的合作，该项目将(1)确定酗酒等有问题的人类行为的风险因素；以及(2)定位由于帕金森氏症或中风等神经疾病而导致的步态异常的最具区分性的感觉特征。

Discovering latent subgroups in a sample is an important problem in many scientific disciplines. Social scientists identify subgroups within a population based on behavioral patterns to examine differential effects of social status. Engineers recognize malfunctions of a manufacturing system based on performance measures to detect design defects. Physicians define subtypes of a disorder on the basis of clinical symptoms to identify associated genetic risk factors. This kind of problem involves two sets of variables: a set of descriptors describing the issue (e.g., behavioral patterns, or symptoms) and a set of moderators or predictors (e.g., social status, or genetic factors). The ability to accurately predict the latent classes (e.g., disease subtypes) from predictors (e.g., genetic risk) in the absence of observed descriptors (e.g., before symptoms are developed) will advance many of these disciplines. This project aims to develop an effective and efficient platform of machine learning algorithms to solve this problem. The team will effectively integrate research and teaching to engage students into the proposed study. Validated methods and software will be broadly disseminated through the project web repository and scientific presentations.This project addresses the latent class discovery and prediction problem by deriving novel and efficient approaches, including multi-view co-clustering, multi-view subspace clustering, multi-objective optimization of co-training, and multi-modal deep learning methods. Parallel and distributed algorithms will be developed to implement and scale up these methods. A streamlined analytics platform will be constructed to maximize the utility of the proposed approaches in real-world applications. The proposed solutions will be evaluated in the analysis of large-scale sensory and behavioral data. By collaborating with domain experts, the project will (1) identify risk factors for problematic human behaviors such as binge drinking; and (2) locate the sensory features most discriminative of gait abnormalities due to neurological disorders such as Parkinson's disease or stroke.