Enhancing Knowledge Tracing through Leakage-Free and Recency-Aware Embeddings
The episode discusses a paper enhancing knowledge tracing by addressing data leakage and incorporating recency-aware embeddings. The authors propose a 'MASK' label to prevent models from cheating by copying answers between related questions, and introduce recency encoding using learnable Fourier features to model student forgetting. This approach improves model accuracy significantly on benchmarks like Duolingo.