Dual-Channel Adaptive Scale Hypergraph Encoders With Cross-View Contrastive Learning for Knowledge Tracing

Area of Science:

• Artificial Intelligence
• Educational Technology
• Machine Learning

Background:

• Knowledge tracing (KT) is crucial for intelligent tutoring systems, predicting learner performance from historical responses.
• Existing deep learning methods (RNNs, attention, GNNs) capture pairwise relationships but overlook higher-order interactions.
• Current models struggle to represent multigranularity knowledge states due to single-channel encoding.

Purpose of the Study:

• To propose a novel KT model, HyperKT, addressing limitations of existing approaches.
• To incorporate non-pairwise, higher-order response information and multigranularity knowledge states.
• To enhance predictive accuracy in intelligent tutoring systems.

Main Methods:

• Developed an adaptive scale hyperedge distillation for knowledge-aware and pattern-aware hyperedges.
• Proposed dual-channel hypergraph encoders (simplified and collaborative hypergraph convolution networks) for global and local state capture.
• Implemented cross-view contrastive learning among hypergraph and line graph views to strengthen supervision.

Main Results:

• HyperKT effectively captures non-pairwise, higher-order features among learner responses.
• The dual-channel encoders successfully represent multigranularity knowledge states.
• Experiments on three real-world datasets show HyperKT outperforms state-of-the-art methods.

Conclusions:

• HyperKT offers a significant advancement in knowledge tracing by modeling complex relationships.
• The proposed methods enhance the representation of learner knowledge states.
• HyperKT demonstrates superior performance, paving the way for more effective intelligent tutoring systems.