基于交互序列特征相关性的可解释知识追踪

Interpretable knowledge tracing based on the feature relevance of interaction sequence

为提高知识追踪(knowledge tracing,KT)模型的可解释性,提出适用于KT事后可解释性的Shapley Value和ISP算法以及可解释性评价指标和谐度,以KT领域经典的深度学习模型DKT为例,计算历史交互与预测结果之间的相关性分数,解释DKT的预测结果。Shapley Value算法计算每次交互对预测结果的贡献,将贡献视为相关性分数;ISP算法基于原序列和模型自身的推理能力构造伪标签,实现对原序列的扰动,计算相关性分数;基于解释方法计算出的相关性分数,使用和谐度指标评价各方法的解释效果。在试验层面,5个公开数据集上的试验结果表明,相对于最优的基线方法,本研究提出的方法取得显著的可解释性效果提升;在具体应用层面,利用可解释性挖掘知识点之间的偏序关系,帮助学生探究更加合理的学习顺序。

In order to improve the interpretability of the knowledge tracing(KT) model,the Shapley Value and ISP algorithms suitable for the post-hoc interpretability of KT and the metrics harmony for the interpretability evaluation were proposed.Taking DKT,a classic deep learning model in the field of KT,as an example,the correlation score between history interaction and prediction results were calculated,explaining the prediction results of DKT.The Shapley Value algorithm calculated the contribution of each interaction to the prediction result,and regarded the contribution as a correlation score;the ISP algorithm constructed pseudo-labels based on the original sequence and the inferring ability of the model,realized the perturbation of the original sequence,and calculated the correlation score;based on correlation scores calculated by different explanation methods,harmony was proposed to evaluate the interpretability of each method.At the experimental aspect,the experimental results on five public datasets showed that,compared with the optimal baseline method,the method proposed in this study achieved a significant improvement in interpretability;at the specific application aspect,interpretability was used to mine the partial order relationship between skills,so as to explore a more reasonable learning sequence for students.