Abstract: With the advancement of educational informatization and the growing emphasis on personalized learning,how to construct accurate and dynamic learner profiles from massive educational data has become a crucial research topic in the field of education.However,existing approaches to learner profiling often suffer from insufficient feature representation,inadequate conside-ration of temporal dynamics,and inconsistencies in tag dependencies.To address these challenges,this paper proposes a method for constructing and predicting temporally dynamic tags.Specifically,the proposed method enhances original feature tags through a multi-source learner feature augmentation strategy,and performs temporal tag modeling under different observation periods based on educational theories.Furthermore,it leverages Teacher Forcing and Scheduled Sampling techniques from natural language processing(NLP) to train a multi-branch Transformer model that deeply captures learner features across varying temporal scales.The proposed approach is evaluated on two publicly available datasets－campus card swipe data and online course records,and the results demonstrate its effectiveness in constructing dynamic temporal tags and learner profiles.

Key words: Learner profile, Feature expansion, Temporal tag, Profile tag construction, Profile tag prediction