Breaking Knowledge Boundaries: Cognitive Distillation-enhanced Cross-Behavior Course Rec Model

The speaker introduces a cross-behavior course recommendation model that fuses cognitive diagnosis with recommendation via a multi-stage knowledge distillation framework. The system builds graphs of course learning, exercise responses, and knowledge inclusion, and encodes learner representations at exercise- and concept-level granularities. Item Response Theory parameters model study factors, which feed a temporal cross-behavior attention module and a recommendation decoder. A triple-stage distillation aligns divergent objectives and mitigates data imbalance. Experiments on MuQubeX and EdNet show consistent gains and ablation validates each component. A case study demonstrates relevance to a learner’s weaknesses, and future work targets noise reduction using richer behavioral signals.