Research map / Graph learning / Knowledge and temporal graphs
Collaborative Cognitive Diagnosis with Disentangled Representation Learning for Learner Modeling
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 Coral, a cognitive diagnosis model that combines a disentangled VAE-style cognitive state encoder with a context-aware iterative neighbor search that builds per-concept collaborative graphs among learners.
Problem
Cognitive diagnosis (predicting learners' knowledge proficiency from practice records) ignores collaborative signals among similar learners and entangles cognitive factors, and explicit learner-similarity connections are not available in education data.
Approach
Disentangled per-concept learner states are learned with a beta-TCVAE-style objective; then, for each concept, K neighbors are selected iteratively by an InfoNCE-based context-aware matching score over relative representations, and a graph network aggregates collaborative states that are residually added to the initial states for prediction.
Key results
- On ASSIST, Coral reaches AUC 74.72 and F1 81.16 with RMSE 43.66, best on F1 and RMSE but below KaNCD on ACC and AUC (KaNCD AUC 75.38).
- On Junyi, Coral achieves ACC 81.15, AUC 80.94, F1 89.12, RMSE 36.08, the best on AUC, F1 and RMSE; KaNCD has the top ACC (81.79).
- On NeurIPS2020EC, Coral achieves the best results on all four metrics (ACC 71.72, AUC 78.88, F1 72.82, RMSE 43.20).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Coral instead of the strongest baseline | Cognitive diagnosis · ASSIST | ACC, AUC, F1-score, RMSE: ACC 72.43 vs 71.53; AUC 75.38 vs 74.72; F1-score 80.60 vs 81.16; RMSE 43.74 vs 43.66 | Mixed |
| Use Coral instead of the strongest baseline | Cognitive diagnosis · Junyi | ACC, AUC, F1-score, RMSE: ACC 81.79 vs 81.15; AUC 80.93 vs 80.94; F1-score 89.02 vs 89.12; RMSE 36.11 vs 36.08 | Mixed |
| Use Coral instead of the strongest baseline | Cognitive diagnosis · NeurIPS2020EC | ACC, AUC, F1-score, RMSE: 71.66 -> 71.72 ACC; 78.57 -> 78.88 AUC; 72.64 -> 72.82 F1-score; 43.21 -> 43.20 RMSE | Helped |
| Add KL regularization (ablation) | Coral · Additional experimental results | ACC: 0.693156 -> 0.709710 | Helped |
| Add KL regularization (ablation) | Coral · Additional experimental results | AUC: 0.659914 -> 0.721823 | Helped |
| Add collaborative learning (ablation) | Coral · Additional experimental results | ACC: 0.667321 -> 0.709710 | Helped |
| Add collaborative learning (ablation) | Coral · Additional experimental results | AUC: 0.606090 -> 0.721823 | Helped |
| Add collaborative learning (ablation) | Coral · Additional experimental results | F1-score: 0.786824 -> 0.803728 | Helped |
| Add collaborative learning (ablation) | Coral · Additional experimental results | RMSE: 0.465231 -> 0.440430 | Helped |
| Replace learned collaborative graph with KNN (ablation) | Coral · Additional experimental results | ACC: 0.708520 vs 0.709710 | No clear effect |
| Replace learned collaborative graph with KNN (ablation) | Coral · Additional experimental results | AUC: 0.721339 vs 0.721823 | No clear effect |
| Replace learned collaborative graph with KNN (ablation) | Coral · Additional experimental results | F1-score: 0.794198 vs 0.803728 | Hurt |
Limitations
- Collaborative neighbor selection is O(M) per learner per step and the paper gives no complexity analysis.
- Reported gains are small on ASSIST and Junyi, where KaNCD matches or beats Coral on some metrics.
- Theoretical EM-equivalence claims are described as approximate in the proofs.
Experimental setup
Datasets: ASSIST Junyi NeurIPS2020EC
Benchmarks: ASSIST Junyi NeurIPS2020EC
Metrics: ACC AUC F1-score RMSE
Baselines: IRT MIRT PMF NCDM KaNCD RCD DCD
Builds on: beta-TCVAE DCD NCDM InfoNCE
Closest papers by meaning
- Multi-Perspective Consolidation Enhanced Cognitive Diagnosis via Conditional Diffusion Model (AAAI 2025 · similarity 0.90)
- DisenGCD: A Meta Multigraph-assisted Disentangled Graph Learning Framework for Cognitive Diagnosis (NeurIPS 2024 · similarity 0.88)
- Enhancing Cognitive Diagnosis Using Un-interacted Exercises: A Collaboration-Aware Mixed Sampling Approach (AAAI 2024 · similarity 0.87)
- Knowledge Is Power: Harnessing Large Language Models for Enhanced Cognitive Diagnosis (AAAI 2025 · similarity 0.87)
- Zero-1-to-3: Domain-Level Zero-Shot Cognitive Diagnosis via One Batch of Early-Bird Students towards Three Diagnostic Objectives (AAAI 2024 · similarity 0.87)
- Disentangling Representations through Multi-task Learning (ICLR 2025 · similarity 0.86)
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.