| With the development of the Internet,more and more online learning platforms have appeard which aim to provide students with rich learning resources through learning platforms to make up for current limitations of traditional education.Konwledge Tracing is a key part of personalized learning which aims to trace students’ dynamic knowledge states based on students’ historical practice records to help platforms provide personalized services,such as recommended propoer questions or personalized learning path generation.However,current knowledge tracing models have some limitations,including limitions on predicting students’ practices and the lack of in-depth exploration of students’ mastery for concepts.Therefore,this paper proposes three knowledge tracing models to optimize limitations of current knowledge tracing models as follow:Factor analysis model captures students’ learning by counting the number of students’ practice on concepts and has strong interpretability and excellent prediction performance.However,factor analysis model usually relies on simple structures to predict,which hardly capture the information contained in data.Therefore,this paper proposes an Input-aware Neural Knowledge Tracing Machine model(INKTM)to improve the prediction performance and enhance the interpretability of model by retaining the second-order feature interactions and introducing the attention mechanism.However,INKTM and other factor analysis models capture students’practice by recording the number of students’ practice on concepts and it is difficult to distinguish the impact of near practice and long-term practice.In addition,the current factor analysis models ignore the rich information contained in the topic itself.Therefore,this paper further proposes a Multi Factors Aware Dual-Attentional model(MF-DAKT).Firstly,to enrich questions’ representations,we design a pre-training method to capture rich questions information,including questions’ relation and difficulty.At the same time,MF-DAKT sets a regularization term of questions’ difficulty to restrict questions’ representations fine-tuning during prediction process.In addition,to highlight the impact of students’ recent practices,we design a recent factor to highlight the impact of recent practices by recording students’ recent exercises in concepts.At the same time,we use a dualattentional mechanism to distinguish the influence of factors and factor interactions.At last,MF-DAKT achieved excellent performance.Then we find that although the knowledge tracing model,including INKTM,and MF-DAKT,have achieved great performance in predicting students’ answers to exercises,they still have limitations in capturing students’ historical exercise sequence information.At the same time,current KT methods tend to be easily over-fitted,caused by the fact that many students have few learning records.Therefore,in this paper,we propose a model called Sequence-Aware Multi-Objective Knowledge Tracing(MO-SKT)to enhance models’ability to model students’exercise sequences and apply multiple losses to help our model learn generalized represen-tations of students with few records.Specifically,we simultaneously extract three kinds of sequence features from students’exercise sequences based on an auto-projector for encoding the number of students’ exercises,a soft-selection mechanism for selecting rele-vant exercises,and a personalized update gate to distinguish the contributions of sequence features to predictions.At the same time,we apply a loss function on sequence features to separately optimize each sequence feature based on students’ responses.Moreover,we jointly design a pair-wise training loss and a disturbance loss to enhance our model’s generalization.We conduct extensive experiments on 4 datasets,and the experimental results validate that MO-SKT can perform better than 13 baselines.At last,we find that due to the lack of labels for students’ mastery of concepts,the previous knowledge tracing methods,including the above three models proposed in this paper INKTM,MF-DAKT,and MO-SKT,only focus on the prediction of students’ exercise,but fail to deeply explore students’ mastery of conepts.In fact,acquiring mastery is significant to personalized learning which can offer interpretable feedback and be beneficial to educational applica-tions.Therefore,in this paper,we aim to acquire students’mastery of concepts via a novel model called Interrelated Matrixes-Based Knowledge Tracing(IMKT)and design a novel metric to estimate the acquired mastery values.Specifically,we first make the best of students’ exercise sequences to trace students’ knowledge states.Then,we construct interrelated representation matrixes of concepts based on the relations between students’mastery and responses to exercises to make acquiring students’ mastery can be interpretable.At the same time,we design to directly apply a loss function on unlabeled mastery combined with a loss function on students’ responses to optimize our model jointly.At last,we conduct extensive experiments on five public datasets.Results validate that IMKT can perform better than 15 baselines on the novel metric and several commonly used metrics. |
