Knowing What Students Know: Using Educational Data Mining to Predict Robust STEM Learning

学生对科学、技术、工程和数学(STEM)概念的理解经常受到学生将肤浅的问题解决策略应用于STEM内容领域的影响。浅层问题解决技能是指适用于特定问题实例，但不适用于类似问题的技能。这项研究项目将研究如何在基于计算机的学习环境中支持对STEM概念的深入理解-特别是遗传学问题的解决。学习环境将支持遗传学过程--建模理解和遗传学溯因推理--即从经验数据到可能产生该数据的遗传过程的推理。过程建模和溯因推理模型都与许多STEM领域相关。开发基于计算机的活动，支持基因组分析的学习，并了解学生对这一主题的可靠知识，有望通过改进STEM学习产生巨大的社会效益。研究将结合认知建模和教育数据挖掘来开发和评估学生在解决问题过程中学习深度的多组件模型。这项拟议的研究解决了在STEM领域观察到的学生问题解决方面的一个众所周知的缺陷：一些学生发展了浅层次的问题解决知识，他们很难在新的情况下应用这些知识。这项拟议的研究将开发一个多成分但简约的学生知识深度模型，该模型预测在直接将知识转移到问题解决课程中连续主题的过程中以及未来学习过程中的个体差异。这项研究将发生在现代遗传学问题解决领域。遗传学是生物学的一个基本、统一的主题，被学生和教师视为生物学中最具挑战性的主题之一，部分原因是它严重依赖于解决问题。遗传学问题的解决依赖于两种类型的知识：遗传过程建模知识和溯因推理技能-从经验数据向后推理到产生数据的遗传过程。拟议的研究将侧重于学生对基因组路径分析的理解，这是一个将经典孟德尔遗传学和分子遗传学联系起来的基础性课题，是理解21世纪生物学进步的关键。本项目将模拟学生遗传学知识的四个组成部分：(1)基因作用的口头陈述性知识；(2)遗传学过程建模知识；(3)溯因推理技能；(4)元认知技能。这项研究将检验关于如何将这些测量整合到一个统一的模型中，以预测未来学习过程中的个体差异；即，学生知识的哪些组成部分影响学生在整个课程中连续主题中学习功能的截取(直接迁移)，以及哪些影响学生学习功能的比率。该项目将采用Genetics Cognitive Tutor，这是一种智能学习环境，为学生提供解决问题所需的循序渐进的支持。该项目利用现有的问题解决模块，将建立一些额外的模块，并将所有模块的附带教学文本放在网上，将学生阅读时间的测量纳入一个健全的学习模式。最近的学生建模工作已经产生了初步的、有希望的结果，为完整的模型奠定了基础：(1)从基本问题解决的逐步准确性推断学生知识的模型可靠地预测了基本的问题解决后测成绩和稳健的学习后测措施(迁移和为未来学习做准备)；(2)学生元认知技能模型更准确地预测了这些稳健的学习措施；(3)自动化模型准确地预测了“浅层”学习。这个项目将把这些组成部分和陈述性知识测量整合到一个全面的框架中，该框架模拟了学生在解决问题时的稳健学习，不仅预测了学生在稳健学习测试中的表现，而且预测了随着学生完成连续的问题解决课程单元，整个学习功能的个体差异。

Student understanding of Science, Technology, Engineering and Mathematics (STEM) concepts frequently suffers from students applying shallow problem solving strategies to STEM content areas. Shallow problem solving skills are skills that apply to a specific instance of a problem but do not generalize to similar problems. This research project will study how deep understanding of STEM concepts can be supported in a computer based learning environment - specifically genetics problem solving. The learning environment will support both genetics process-modeling understanding and genetics abductive reasoning - that is reasoning from empirical data to the genetic processes that could generate that data. Both process modeling and abductive reasoning models are relevant to many STEM domains. Developing computer-based activities that support learning of genomic analysis, and understanding robust student knowledge of this topic promise to yield large societal benefits by improving STEM learning.The research will combine cognitive modeling and educational data mining to develop and evaluate a multi-component model of the depth of student learning during problem solving. The proposed research addresses a well-documented shortcoming in student problem solving, observed across STEM domains: some students develop shallow problem-solving knowledge that they have difficulty applying in new situations. The proposed research will develop a multi-component, yet parsimonious, model of students' depth of knowledge that predicts individual differences in direct transfer of knowledge to, and future learning of, successive topics in a problem-solving curriculum. This research will occur in the domain of modern genetics problem solving. Genetics is a fundamental, unifying theme of biology and is viewed as one of the most challenging topics in biology by students and instructors, in part because it relies heavily on problem solving. Genetics problem solving relies on two types of knowledge: genetic process-modeling knowledge, and abductive reasoning skills - reasoning backwards from empirical data to the genetic processes that generated the data. The proposed research will focus on student understanding of genomic pathway analysis, a foundational topic linking classic Mendelian genetics and molecular genetics that is key to understanding advances in 21st century biology. This project will model 4 components of students' genetics knowledge: (1) verbal declarative knowledge of gene action; (2) genetics process modeling knowledge; (3) abductive reasoning skill; and (4) metacognitive skills. The research will examine hypotheses about how to integrate these measures into a unified model that predicts individual differences in future learning processes; that is, which components of student knowledge affect the intercept (direct transfer), and which affect the rate, of students' learning functions in successive topics across the curriculum. The project will employ the Genetics Cognitive Tutor, an intelligent learning environment that provides students step-by-step support as needed in solving problems. The project leverages existing problem-solving modules, will build some additional modules, and will bring the accompanying instructional text for all modules on-line to incorporate student reading-time measures into a robust learning model. Recent student modeling efforts have yielded preliminary, promising results that lay the foundation for a complete model: (1) models that infer students' knowledge from step-by-step accuracy in basic problem-solving reliably predict both basic problem-solving posttest performance, and robust learning posttest measures (transfer and preparation for future learning; (2) models of students' metacognitive skills more accurately predict these robust learning measures; and (3) models of automaticity accurately predict "shallow" learning. This project will integrate these components and declarative knowledge measures into a comprehensive framework that models students' robust learning in problem-solving and predicts not only time-slices of student performance on robust-learning tests, but predicts individual differences in entire learning functions as students complete successive problem-solving curriculum units.