AMBER
- Type
- Open source
- Year
- 2025
- Status
- active
An adaptive meta-balancing framework for integrating heterogeneous graph signals in knowledge tracing.
AMBER tackles imbalance between heterogeneous graph views used to model student learning. A pretrained dual-graph teacher provides ensemble knowledge while a meta-distillation mechanism adjusts the teacher according to student-model feedback.
The official implementation includes the model, preprocessing utilities, and experiment configuration used for the CIKM 2025 study.