A Novel Transfer Learning Approach for Personalizing Learning Experiences Using Pretrained Transformers

Abstract

The growth in the application of the technology-based learning platforms has resulted in the development of very high demands for intelligent personalized learning solutions that will be able to customize the learning material according to the individual requirements of the learners. Despite this, many of the existing recommender systems in education rely heavily on the usage of the static learner model and conventional machine learning approaches, which cannot take into consideration the dynamic nature of the contexts that influence the behavior of the learners. This paper investigates a novel transfer learning approach for personalized learning using pretrained transformers. In the suggested system, transformers can be used to generate contextual embeddings, adapt learner profiles, and transfer knowledge to enhance the efficiency of educational recommendations and learning results prediction. The new scheme employs pretrained Bidirectional Encoder Representations from Transformers (BERT), which are then fine-tuned using learner interaction datasets consisting of behavioral trends, past academic achievements, discourses, and text feedback. The experimental evaluation was performed using educational datasets consisting of structured and unstructured interaction data of learners. The dataset was split into training, validation, and testing sets based on a stratified data partitioning approach. It can be seen from the experimental results that the suggested approach has shown effectiveness when assessed against different measures. Specifically, the model obtained the accuracy of 96.4%, the precision of 95.8%, the recall of 95.1%, and the F1-score of 95.4%, performing better than collaborative filtering, SVM, CNN, and LSTM models. In terms of recommendation accuracy, the value was equal to 96.9%. Meanwhile, the use of transformer-based transfer learning allowed saving up to 31% in inference time compared to conventional transformer-based algorithms. Learner engagement has increased by 19.3%, while assessment performance has improved by 16.7% after applying personalized recommendations. The results show significant benefits of transfer learning with transformers for educational purposes.