Visual Analysis Of Deep Knowledge Tracing Based On Neuron Behavior

With the continuous integration of computer-assisted learning and artificial intelligence,knowledge tracing(KT)[16]referring to trace the knowledge state of the student based on the history interaction collected from the online educational platform,becomes a hot issue in the domain of intelligent tutoring systems.In the early days,Bayesian Knowledge Tracing(BKT)[16],a knowledge tracing method based on Dynamic Bayesian Network,has been proposed and following such seminal work,the extensions of BKT have been developed continually and employed to KT broadly.Then,with the rapid development of deep learn-ing,Deep Knowledge Tracing(DKT)[44],based on recurrent neural network[25],has been suggested and achieves better performance than BKT,which receives increasing attention.However,comparing to BKT,DKT is lack of specific encoding of human domain knowledge due to the inherent property of the black box of deep learning neural network,which leads to the insufficiency of interpretability of DKT.In the domain of education,it is crucial to make it clear that the logical and reasonable relationship among the knowledge components because the lucid explanation of the knowledge components facilitates to diagnose the stu-dent's mastery level according to the estimate of the student's performance.Furthermore,the improvement of the interpretability of DKT not only contributes to promotes its employment in the domain of education but also helps to optimize the architecture of the model based on the understanding of the working mechanism.Given this challenge,the thesis aims to ex-plore the working mechanism of the neurons within DKT through visual analysis to discover the human-understandable representation.Based on DKT and the visualization tools,the thesis proposes the hypothesis that the hidden representation can describe the relationship among knowledge components or KCs.The inspirations of the hypothesis are the speculation of the context of KT and the related experiment conducted by the author of DKT,which supports that DKT is capable of discov-ering the dependency between knowledge components.To validate the proposed hypothesis,LSTMVis,a visualization tool for RNN,developed by the visual computing group at Har-vard,is adopted to implement visual analysis on DKT based on the research hypothesis and the experiment.Firstly,to be directed against the hypothesis,Apriori,an algorithm for dig-ging association rules,is adopted to extracts the sets of knowledge components that occur frequently.Secondly,according to the result of the dataset analysis,four sets of input se-quences are designed for targeted visual analysis.And the result of visual analysis suggests that some neurons within DKT are activated to detect the association information of knowl-edge components,hierarchy information of knowledge components,and similar practicing context,respectively.Moreover,based on LSTMVis,an auxiliary deep learning visualiza-tion tool is developed to probe the hidden states against non-target labels and the overall comparison of the precise hidden representation of the selected time steps.Last,the thesis applies this tool to DKT for visual analysis.And the result shows that DKT works as co-ordinating most neurons,including the inactivated neurons,to have the same pattern when detecting the association information of knowledge components,hierarchy information of knowledge components,and similar practicing context,respectively.