Research map / Probabilistic and scientific modeling / Causal discovery
Causal-Driven Skill Prerequisite Structure Discovery
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CSPS, a two-stage causal-driven method that recovers latent skill covariance from student exercise performance via a linear-Gaussian latent variable model, then learns skill prerequisite structure with NB-Search skeleton search and DS-SPV direction search.
Problem
Discovering skill prerequisite structure from student performance data when skill mastery is latent, exercises load on multiple skills, and response logs are sparse/missing.
Approach
Model skills as latent variables and exercises as observed variables in a linear-Gaussian latent variable model, derive the posterior skill covariance matrix, then apply causal-discovery-style local search (conditional independence tests via Fisher's z on the covariance, plus standardized partial variance log-ratios) to get the skeleton and orient edges.
Key results
- On FrcSub, CSPS achieves TPAR of 0.84 and the best adjacency F1-AR, with F1-AR about 0.80 and F1-OR about 0.60.
- On simulated Sync1 with 500 and 2000 samples, CSPS achieves the highest F1-AR of 1.000 among compared methods.
- On Alg0506, CSPS reaches about 70% F1-AR and 45% F1-OR, while CITS performs near zero TPAR.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CSPS instead of the strongest baseline | Skill prerequisite discovery · Sync1, D=500 | F1-AR: 1.000±.000 -> 1.000±.000 | No clear effect |
| Use CSPS instead of the strongest baseline | Skill prerequisite discovery · Sync2, D=500 | F1-AR: 0.667±.000 -> 0.750±.000 | Helped |
| Use CSPS instead of the strongest baseline | Skill prerequisite discovery · Sync2, D=150 | F1-AR: 1.000±.000 -> 0.500±.163 | Hurt |
| Use CSPS instead of the strongest baseline | Skill prerequisite discovery · Sync1, D=1000 | F1-OR: 0.813±.228 -> 0.900±.061 | Helped |
| Use CSPS instead of the strongest baseline | Skill prerequisite discovery · Sync2, D=1000 | F1-OR: 0.462±.000 -> 0.700±.105 | Helped |
| Use CSPS instead of the strongest baseline | Skill prerequisite discovery · Sync1, D=2000 | F1-OR: 1.000±.000 -> 0.667±.000 | Hurt |
| Use CSPS instead of the strongest baseline | Skill prerequisite discovery · Sync2, D=2000 | F1-OR: 0.667±.000 -> 0.500±.000 | Hurt |
Surprising findings
- CITS, a prior causal-discovery approach, nearly fails on real-world FrcSub and Alg0506 data (TPAR near zero).
- Orientation is consistently harder than adjacency recovery across all methods.
Limitations
- Relies on Gaussian assumptions and Fisher's z-test on an estimated latent covariance treated as a sample correlation with an assumed sample size D, lacking statistical guarantees.
- Performance on real data is modest (F1-OR about 0.45-0.60), with noise in logs and mislabeled skills hurting results.
- Baselines were limited to discrete-variable methods for simulated data and only CITS could be applied to continuous data.
Experimental setup
Datasets: Sync1 (simulated) Sync2 (simulated) FrcSub Alg0506
Benchmarks: Sync1 Sync2 FrcSub Alg0506
Metrics: F1-AR (adjacency) F1-OR (orientation) TPAR TDAR TPOR TDOR
Baselines: EPS-ND CITS PARM CMPD BE-NMC DIDACT
Builds on: PC algorithm Markov blanket local discovery (MMHC) standardized partial variance (SPV) Brunskill EPS-ND CITS
Closest papers by meaning
- K12-KGraph: A Curriculum-Aligned Knowledge Graph for Benchmarking and Training Educational LLMs (NeurIPS 2026 · similarity 0.84)
- Predictive, scalable and interpretable knowledge tracing on structured domains (ICLR 2024 · similarity 0.83)
- Local Causal Structure Learning in the Presence of Latent Variables (ICML 2024 · similarity 0.82)
- Progressive distillation induces an implicit curriculum (ICLR 2025 · similarity 0.82)
- An Autoencoder-Like Nonnegative Matrix Co-Factorization for Improved Student Cognitive Modeling (NeurIPS 2024 · similarity 0.82)
- Score-Based Causal Discovery of Latent Variable Causal Models (ICML 2024 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.