CRII: III: Modeling Student Knowledge and Improving Performance when Learning from Multiple Types of Materials

随着国民对高等职业教育的兴趣与日俱增，人们对在线学习系统的兴趣也迅速增长。在线学习系统，如大规模开放在线课程和智能辅导系统，旨在通过大规模提供高质量、负担得起和可获得的教育来为社会做出贡献。它们通过为高需求的工作培养熟练的专业人员，对国家繁荣的推进产生了重大影响。要实现如此高影响力的目标，需要自动化工具来帮助我们了解学生的学习过程，并回答一些问题，例如通过观看视频讲座获得了什么知识(领域知识建模)，学生的知识状态是什么(学生知识建模)，以及特定学生在测试中的表现(预测学生表现)。理想情况下，这些工具应该模拟学生从各种学习材料类型(如问题、阅读材料和视频讲座)中进行的学习，并捕捉可分级和不可分级学习资源的组合所提供的知识广度。然而，目前的工具仅限于单一类型的学习材料(通常是“问题”)，忽视了学生可以学习的学习材料的异质性。这个项目旨在通过提出一个综合的研究和教育计划来更好地理解在线教育系统中学生的学习过程：(1)模拟学生与可分级和不可分级学习材料类型的交互；(2)将所提出的模型与学习材料内容相结合；(3)通过使用真实世界的在线教育数据集来对所提出的模型进行评估。该项目将为研究生和本科生提供学习和研究的机会。为了实现在线教育系统中改善学生学习过程的目标，研究人员开发了多视图机器学习算法，在最大化学习数据的多个视图之间的相关性的同时，最小化学生表现预测的误差。在这个项目的第一年，将使用学生的活动序列来构建一个模型，该模型可以捕捉到可分级学习材料和非可分级学习材料之间共享的潜在知识空间。在这个项目的第二年，学习材料的内容信息，包括专家标签，将被纳入学习模型，以改进它。该模型旨在发现内容信息与共享的潜在知识空间之间的关系。使用预测学生表现的任务对项目结果进行评估。这个项目是领域适应、序列建模和教育数据挖掘的交叉点。该模型的灵感来自于典型相关分析，它是一种传递信息和适应学生活动数据的不同视图的方法，同时将学生的学习过程建模为知识获取的序列。这是对学生建模问题的一种新的处理方法，具有序贯领域适应的观点，有助于未来的研究方向，如个性化教育和改善在线学习环境中的学生保留。这项工作提出了新颖的顺序和内容感知的领域自适应和多变量分析模型，该模型将来自多个顺序数据源和时间不变的内容资源的信息同时结合在一起。在学生知识建模任务的激励下，这些模型是通用的，可以应用于包括领域适应问题和推荐系统在内的广泛研究。开发的解决方案将在期刊和会议场地上公布，项目网站将提供对结果的访问，以及将在GitHub上提供的开发和评估模型的代码参考。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。

As the national interest in higher and professional education has been increasing, interest in online learning systems has also grown rapidly. Online learning systems, such as Massive Open Online Courses and Intelligent Tutoring Systems, aim to contribute to the society by providing high quality, affordable, and accessible education, at scale. They highly impact advancement of the national prosperity by preparing skillful professionals for high-demand jobs. Delivering such high-impact goals requires automatic tools that can help us understand students' learning process and answer questions such as what knowledge is gained by watching a video lecture (domain knowledge modeling), what is a student's state of knowledge (student knowledge modeling), and how a specific student would perform on a test (predicting student performance). Ideally, these tools should model student's learning from various learning material types (such as problems, readings, and video lectures) and capture the knowledge span offered by combinations of gradable and non-gradable learning resources. However, the current tools are limited to a single type of learning material (typically, "problems"), ignoring the heterogeneity of learning materials from which students may learn. This project aims to achieve a better understanding of students' learning process in online educational systems by presenting an integrated research and education plan (1) to model student interactions with both gradable and non-gradable learning material types, (2) to integrate the proposed models with learning material content, and (3) to evaluate the proposed models by experimenting with real-world online educational datasets. The project will provide learning and research opportunities to graduate and undergraduate students.To achieve the goal of improving students' learning process in online educational systems, the researchers develop multi-view machine learning algorithms that minimize the error of student performance prediction while maximizing the correlations among multiple views of the learning data. In the first year of this project, using activity sequences of students a model will be built that can capture a shared latent knowledge space among sets of gradable learning material and non-gradable ones. During the second year of this project, content information of learning materials, including expert labels, will be included in the learning model in order to improve it. This model aims to discover the relationship between content information and the shared latent knowledge space. The project results are evaluated using the task of predicting student performance. This project is at the intersection of domain adaptation, sequence modeling, and educational data mining. The model is inspired by Canonical Correlation Analysis as an approach for transferring information and adapting various views to student activity data, while modeling student learning process as a sequence of knowledge acquisitions. This is a novel treatment of the student modeling problem, with a sequential domain-adaptation view, that facilitates future research directions, such as personalized education and improved student retention in online learning environments. This work contributes novel sequential and content-aware domain adaptation and multi-variate analysis models that combine information from multiple sequential data sources and time-invariant content resources at the same time. While motivated by the task of student knowledge modeling, the models are general and can be applied to a broad spectrum of research including domain adaptation problems and recommender systems. The developed solutions will be presented in journals and conference venues, and the project website will provide access to the results, with references to code for the developed and evaluated models that will be available at GitHub.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.