Transfer Learning for Course Recommendation Systems: A Comparative Study of Caser and GRU4Rec Models

A comparative analysis of transfer learning approaches in course recommendation systems is conducted using two deep learning architectures: Convolutional Sequence Embedding Recommendation (Caser) and Gated Recurrent Unit for Recommendation (GRU4Rec). The study evaluates the effectiveness of pre-training these models on a large-scale educational dataset from edX and fine-tuning them on the Coursera and ThaiRobotics datasets, which consist of learning sequences from students at King Mongkut's University of Technology Thonburi (KMUTT) in Thailand, focusing on engineering and programming courses. Experimental results demonstrate significant performance improvements through transfer learning. Caser showed an improvement of up to 145.4% in precision and 84.4% in recall on the Coursera dataset, while GRU4Rec achieved enhancements of 302.0% and 120.3%, respectively. The Caser model excelled in capturing short-term dependencies and prerequisite relationships within engineering courses, showing a 65.7% improvement in hit rate on Coursera and 12.0% on ThaiRobotics after fine-tuning. Additionally, GRU4Rec exhibited complementary strengths by effectively modeling long-term progression, achieving a 123.1% improvement in hit rate on Coursera and 7.7% on ThaiRobotics. These findings validate the effectiveness of transfer learning in enhancing recommendation accuracy and learning efficiency, particularly in scenarios with limited target domain data. Moreover, they highlight the distinct architectural advantages of different models for various aspects of educational sequence modeling.