Doctoral Dissertation Research: Bringing the Power of Deep Learning to Large-Scale Ordinal Data Classification

本博士论文研究项目将为有序数据开发一种更有效、更高效的分类方法。由于数字数据的使用呈爆炸式增长，拥有数十万甚至数百万条记录的有序数据集的数量急剧增加。例如，在亚马逊和Yelp等网站上发现的评级调查，大公司客户满意度/网络推广者调查，以及病历记录的汇总。现有的分类方法不足以分析大型有序数据集。研究人员将开发一种分类方法，以促进对一系列应用领域的大型有序数据集的分析。将开发开放源码软件并向公众开放。将提供学习材料和案例研究，供教育工作者在相关课程和体验式学习机会中采用。作为博士论文研究改进奖，支持有前途的学生建立强大的、独立的研究生涯。这项博士论文研究将开发一种高度可扩展的顺序分类方法，既可以应用于结构化和非结构化(例如图像和文本)顺序数据。该方法的一个核心部分是称为有序超平面损失(OHPL)的损失函数。OHPL是专门针对具有序数类的数据而设计的，它使得深度学习技术能够应用于序数分类问题。通过最小化OHPL，深度神经网络学习将数据映射到一个最优空间，在该空间中，点与其类质心超平面之间的距离最小化，同时保持类之间的非平凡顺序关系。初步实验结果表明，基于OHPL优化的深度神经网络在多个数据集上的分类精度明显优于现有的分类方法。这项研究将考察将基于OHPL的学习扩展到大序数数据的策略。研究人员将把基于OHPL的学习应用于现实生活中的关键应用，如确定疾病的严重程度/阶段。他们将开发一个关于OHPL深度学习战略的现成开放源码包并将其公开提供。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

This doctoral dissertation research project will develop a more effective and efficient classification method for ordinal data. Due to the explosion in the use of digital data, there has been a dramatic increase in the number of ordinal datasets that have hundreds of thousands or even millions of records. Examples include ratings surveys found on sites like Amazon and Yelp, large corporation customer satisfaction/net promoter surveys, and the aggregation of medical history records. Current classification methods are inadequate for analyzing large ordinal datasets. The investigators will develop a classification method that facilitates the analysis of large ordinal datasets across a spectrum of application domains. Open-source software will be developed and made publicly available. Learning material and case studies will be made available for educators to adopt in relevant courses and experiential learning opportunities. As a Doctoral Dissertation Research Improvement award, support is provided to enable a promising student to establish a strong, independent research career.This doctoral dissertation research will develop a highly scalable ordinal classification method that can be applied to both structured and unstructured (e.g., images and text) ordinal data. A core component of the method is a loss function that called Ordinal Hyperplane Loss (OHPL). OHPL is particularly designed for data with ordinal classes and enables deep learning techniques to be applied to the ordinal classification problems. By minimizing OHPL, a deep neural network learns to map data to an optimal space where the distance between points and their class centroid hyper-plane are minimized while a nontrivial ordinal relationship among classes are maintained. Preliminary experimental results indicate that deep neural network with OHPL optimizing significantly outperforms the state-of-the-art alternatives on classification accuracies across multiple datasets. This research will examine strategies that scale the OHPL based learning to big ordinal data. The investigators will apply the OHPL-based learning to real-life critical applications such as determining the severity/stages of a disease. They will develop a ready-to-use open-source package on the OHPL deep learning strategy and make it publicly available.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.