Learning path recommendation with multi-behavior user modeling and cascading deep Q networks

Abstract

An online learning platform has become an important channel for learners to obtain knowledge due to its easy access and rich resources. In order to meet online learners' short-term needs with frequent changes and long-term learning goals during learning process, this paper focuses on user modeling and learning path recommendation, and we propose a new method for learning path recommendation through multi-behavior user modeling and cascading deep Q networks (cDQN-PathRec). Our model uses a knowledge graph-based multi-behavior transformer architecture for users’ state modeling, in which a learner's knowledge background, learning styles, learning settings, and learning preferences are taken into consideration. We use a cascading DQN with a two-level reward function to help an agent converge towards a balanced overall and local optima and to generate a learning path recommendation. Comprehensive experiments on two real-world online learning datasets demonstrate effectiveness of the proposed cDQN-PathRec method.