摘要
认知追踪是一种动态的学习主体建模技术,已被广泛应用于智能教育领域。随着以深度学习为代表的新一代人工智能技术的迅猛发展,基于深度学习的认知追踪——深度认知追踪成为当前智能教育领域的研究热点。针对现有的深度认知追踪方法普遍只利用单个时间步信息来引导模型拟合学习者的行为数据,容易导致因模型监督信号不足而难以挖掘复杂行为背后蕴含的稳定性因素这一问题,文章提出一种基于多阶拟合机制的深度认知追踪方法。该方法通过融合多个邻接时间步的信息来增强模型监督信号,并设计了相应的权重规则和去重规则,以减少远距离、冗余性信息带来的干扰。通过在4个基准数据集上的对比实验与预测结果热力图的可视化分析,文章发现该方法不仅有效提升了模型的预测性能,而且赋予了模型更强的可解释性。
Knowledge Tracing(KT)is a dynamic learner modeling technology,which is widely used in the field of intelligent education.With the rapid development of the new generation of artificial intelligence technology represented by deep learning,Deep Knowledge Tracing(DKT)has become a current research hotspot in the field of intelligent education.In view of the situation that existing DKT methods generally use only a single time step information to guide the model to fit learners’behavior data,which easily leads to the problem that it is difficult to mine the stability factors behind complex behaviors due to insufficient model supervision signals,this paper proposed a DKT method based on multi-step fitting mechanism.This method enhanced the model supervision signals by fusing the information of multiple adjacent time steps,and designed the corresponding weight rules and de-duplication rules to reduce the interference caused by long distance and redundant information.Through comparative experiments of four benchmark data sets and visual analysis of heat map of predicted results,this paper found that the method not only effectively improved the prediction performance of the model,but also endowed the model with stronger interpretability.
作者
孙建文
刘盛英杰
刘三女牙
张慧芳
李卿
SUN Jian-wen;LIU Sheng-ying-jie;LIU San-nyu-ya;ZHANG Hui-fang;LI Qing(Faculty of Artificial Intelligence in Education,Central China Normal University,Wuhan,Hubei,China 430079)
出处
《现代教育技术》
CSSCI
2021年第10期103-109,共7页
Modern Educational Technology
基金
教育部人文社会科学研究青年基金项目“面向启发式教学的智能课堂编排模型与方法研究”(项目编号:20YJC880083)
国家自然科学基金青年项目“课堂环境下基于多传感信息的学习注意力识别研究”(项目编号:61807012)
华中师范大学中央高校基本科研业务费项目“智能课堂共享调节学习多模态感知与融合计算研究”(项目编号:CCNU20ZN007)资助。
关键词
认知追踪
深度学习
深度认知追踪
目标函数
多阶拟合
knowledge tracing
deep learning
deep knowledge tracing
objective function
multi-step fitting
作者简介
孙建文,副教授,博士,研究方向为智能导学、智慧课堂,邮箱为sunjw@ccnu.edu.cn;通讯作者:李卿。
