Probabilistic Embeddings Enable Explainable Knowledge Tracing
The recently introduced framework known as Probabilistic Logical Knowledge Tracing (PLKT) seeks to enhance the interpretability of knowledge tracing by substituting deterministic vector embeddings with Beta-distributed probabilistic embeddings. Knowledge tracing evaluates student knowledge states based on learning interactions to forecast future performance. Although deep learning models have increased accuracy, they depend on unclear latent state transitions. PLKT redefines prediction as a goal-conditioned reasoning process that utilizes historical behaviors, incorporating clear logical operations such as conjunction for better transparency. This method tackles the interpretability challenges faced by current knowledge tracing models.
Key facts
- PLKT uses Beta-distributed probabilistic embeddings instead of deterministic vectors.
- The framework models uncertainty of historical behaviors.
- Explicit logical operations (e.g., conjunction) are used for reasoning.
- Knowledge tracing predicts student performance based on learning interactions.
- Deep learning KT models have high accuracy but limited interpretability.
- PLKT formulates prediction as goal-conditioned evidence reasoning.
Entities
—