Sledování znalostí (KT)

  • Sledování znalostí (KT)

Properties1
tagsstrand-technology

🚧 Stránka je ve výstavbě.

KT bude jednou z klíčových komponent Sovy, ale je třeba ho dobře navrhnout. Zatím sem sbírám jenom zajímavé akademické práce.

Hoa, Le, TBD

Summary. This study investigates whether intermediate step-level evidence provides additional predictive information in Bayesian Knowledge Tracing beyond a single terminal correctness outcome. Using the ASSISTments 2009–2010 dataset, the authors constructed an evaluation cohort of 8,995 instances from 1,494 learners across 33 skills. They found that integrating observed scaffold-step responses improves the discrimination of subsequent performance compared to collapsing a multi-step learning episode into one binary response.

Why it matters. When building knowledge tracing models, treating a complex problem as a simple right-or-wrong interaction discards meaningful information about the learner’s problem-solving process. Incorporating the steps a student takes provides a more accurate picture of their actual understanding. This is crucial for effective knowledge tracing.

Connects to. knowledge_tracing

Truong

Summary. This study introduces a traceable belief-governance layer for prerequisite knowledge graphs in tutoring systems, using data from 247,606 learners and 722 concepts. The system represents crowd-rated prerequisite relations as graded strengths and separates majority denial from disagreement about strength. It retains the ratings behind admitted relations and routes non-admitted candidates to a review log, combining them with an expert curriculum map.

Why it matters. Building knowledge graphs from crowd ratings requires a transparent governance layer that preserves the rationale behind prerequisite links.

Connects to. knowledge_tracing

Muerers

KT can be done outside of STEM fields.

Takeaway: if you don’t have large datasets, DKT (neural model) can be outperformed by a logistic regression model (AFM, PFA). Deep models can outperform nuanced predictions but lack interpretability. Logistic models can be traced down to one parameter.

Zdroje