Research On Knowledge Tracing Based On User Stratification And Transformer

With the rapid development of educational information technology,the adaptive learning platform based on deep learning plays an important role in improving learning efficiency by aggregating educational resources and providing personalized teaching,and has become a research hotspot in the field of education.Adaptive learning is an educational method which uses artificial intelligence to orchestrate the interaction with the learners and provide personalized teaching resources to meet the unique learning needs of each learner.The core issue is how to build personalized learning sequence for learners in the massive learning resources.Knowledge tracing algorithm is used in the field of adaptive learning by evaluating learners’ knowledge levels and predicting learners’ next learning performance.Although some deep knowledge tracing algorithms have made a great success,there are still some shortcomings.On the one hand,the input sequence length of the current deep knowledge tracing model is fixed and limited.It is difficult to process long sequence features and ignore the different features of student groups at different levels.On the other hand,the persistent problems of deep learning are poor interpretability and lack of learned features.In addition,most adaptive learning sequences rely on the zones of to proximal development recommend questions with a stable correct rate range for learners,ignoring the degree to which different items can improve learners’ abilities.Aiming at the above problems,this paper explored some work.Firstly,this paper proposed a knowledge tracing model based on user stratification and transformer.It uses the Item Response Theory model to rank the performance of users,and then trains through models of different layers,so that it could better consider the different characters of different learner groups.Based on the IRT model of mathematical statistics,the interpretability of the model has been improved.In addition,to cope with the lack of learning features,this paper improves the transformer structure and encodes the input item label data,answering process and result data through two multi-head attention networks to fully mine the hidden features in the dataset.This article conducts experiments on three data sets of Ed Net,ASSIST2017 and Byte Dance Pony2021,which have achieved good performance in ACC and AUC indicators.Finally,in order to improve the building process of the adaptive learning sequence,this paper designs an exercise recommendation system based on the growth score,and builds the learning resource sequence with the maximum profit by predicting accuracy rate and calculating the improvement value of ability.It also verified the applicability of the proposed model.