从数据视角透析认知追踪:框架、问题及启示

Knowledge Tracing from the Data Perspective:Conceptual Framework,Key Issues,and Educational Enlightenment

认知追踪是一种数据驱动的学习者建模技术,被广泛应用于智能导学、智能课堂编排等系统。尤其是2015年深度神经网络被引入认知追踪任务以来,认知追踪成为智能教育领域的研究热点。针对当前研究普遍存在的“重模型、轻数据”以及数据处理不一致等问题,本研究基于近六年国内外35篇有关认知追踪的论文,全面梳理和分析其中被高频使用的数据集,提出以学生、知识、问题三个对象及六类交互关系为核心的认知追踪概念框架,为深层次理解数据内涵和统一数据操作提供指导。本研究还运用该框架对数据集特征进行分类,围绕数据重复、数据顺序、支架题目、技能缺失以及多技能题目等关键问题进行数据一致性分析,特别是针对多技能题目,提出了基于多热编码的表示方法。本研究最后从五方面讨论了认知追踪及未来智能教育的发展趋势:从个体自主学习到多模式混合学习、从单一学习行为到多模态数据融合、从深度学习算法黑箱到可解释分析、从数据驱动到数据与知识联合驱动,以及从技术意识垄断回归教育价值本位,为拓展认知追踪研究边界、促进智能教育创新突破提供参考与指引。

Knowledge tracing is a data-driven intelligent cognitive modeling technology,which is widely used in intelligent tutoring systems,smart classroom orchestration,and other systems.Since the first application of deep neural networks in the knowledge tracing task in 2015,knowledge tracing has rapidly become a research hotspot in intelligent education.Aiming at the common problems of“emphasizing model,neglecting data”and inconsistent data processing in current research,this paper systematically investigates 35 papers on knowledge tracing at home and abroad in recent six years,comprehensively analyzes the data sets used frequently,and proposes a conceptual framework based on the three objects of“Student&Knowledge&Problem”and six types of interactions,so as to guide a deep understanding of data connotation and related operations.Then,the framework is used to classify the features of three datasets and conduct data consistency analysis and standardization processing around key issues such as data duplication,data sorting,scaffolding problem,skill deficiency,and multi-skill question encoding.Especially for multi-skill questions,a multi-skill encoding method based on multi-hot is proposed.Finally,the development trend of knowledge tracing and intelligent education in the future is discussed from five aspects:from individual autonomous learning to mixed learning,from learning behavior to multimodal data fusion,from deep learning algorithm black box to interpretable analysis,from data-driven to data-knowledge joint driving,and from technology consciousness monopoly to education value,which provides reference and guidance for expanding the research boundary of knowledge tracing and promoting the innovation breakthrough of intelligent education.