融合因果推断的动态可解释知识追踪模型  

Dynamic interpretable knowledge tracing model incorporating causal inference

作　　者：鲁法明 王卓凡 包云霞 王晓亮 LU Faming;WANG Zhuofan;BAO Yunxia;WANG Xiaoliang(College of Computer Science and Engineering,Shandong University of Science and Technology,Qingdao 266590,China;College of Mathematics and Systems science,Shandong University of Science and Technology,Qingdao 266590,China)

机构地区：[1]山东科技大学计算机科学与工程学院,山东青岛266590 [2]山东科技大学数学与系统科学学院,山东青岛266590

出　　处：《山东科技大学学报（自然科学版）》2024年第4期111-120,共10页Journal of Shandong University of Science and Technology(Natural Science)

基　　金：新一代人工智能国家科技重大专项项目(2022ZD0119501);国家自然科学基金项目(52374221);山东省自然科学基金项目(ZR2022MF288,ZR2023MF097);山东省泰山学者特聘专家支持计划项目(ts20190936);青岛西海岸新区科技计划专项项目(202209)。

摘　　要：可解释知识追踪对于教学诊断和优化具有重要意义。目前代表性的可解释知识追踪模型缺乏对知识点间因果关系的深入考量,同时未能关注模型中存在的特征混淆问题,影响预测性能。针对上述问题,提出一种融合因果推断的动态可解释知识追踪模型。首先利用因果推断算法挖掘知识点间掌握程度上的因果关系,并制定规则辨别因果效应强度,挖掘得到知识点因果效应特征图,并从中提取知识点外延影响因子特征,作为学生答题正确性预测的特征之一;其次,基于领域知识构建学习能力、习题难度和作答正确率之间的结构因果模型,采用后门调整的方法去除混杂因子的影响;然后,以迭代的方式进行学习能力和答题偏好特征的动态更新;最后,借助树增广朴素贝叶斯分类器实现可解释性知识追踪。在多个公开的数据集上进行实验验证表明,所提模型在保证可解释性的同时可提高预测准确性。Interpretable knowledge tracing is important for teaching diagnosis and optimization.The existing representative interpretable knowledge tracing models lack in-depth consideration of causal relationships among concepts and fail to pay attention to possible feature confusion in the models,which in turn affects the prediction performance.To address the above problems,a dynamic interpretable knowledge tracing model incorporating causal inference was proposed.First,the causal inference algorithm was used to excavate the causal relationships among concepts in terms of mastery level,and rules were formulated to identify the intensity of causal effects and excavate the concept causal effect feature graph.The features of the knowledge points’extraneous influencing factors were also extracted from it and were used as one of the features for predicting the correctness of students’answers.Second,based on the domain knowledge among the learning ability,the difficulty of the problems,and the correctness rate of the answers,a structural causal model was constructed and the backdoor adjustment method was used to remove the influence of confounding factors.Then,the dynamic updating of learning ability and problem-answer preference features was carried out in an iterative manner.Finally,interpretable knowledge tracing was realized by means of the tree augmented naive Bayes classifier.Experimental validation on several publicly available datasets shows that the proposed model can improve prediction accuracy while ensuring interpretability.

关 键 词：可解释知识追踪 因果推断 树增广朴素贝叶斯分类器 答题正确性预测 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]