Research map / Graph learning / Knowledge and temporal graphs
DisenGCD: A Meta Multigraph-assisted Disentangled Graph Learning Framework for Cognitive Diagnosis
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 DisenGCD, which disentangles cognitive diagnosis into three graphs (student-exercise-concept interaction, exercise-concept relation, concept dependency) and learns student representations with a meta multigraph module while using GAT for exercises and concepts.
Problem
Graph-based cognitive diagnosis learns exercise and concept representations on a unified interaction graph, making them sensitive to noise in student responses; shallow exercise representations are also unused for student learning.
Approach
Disentangle the interaction graph into relation and dependency graphs so exercise and concept representations are learned independently of student interactions, and learn student representations with a meta multigraph aggregator with learnable propagation paths that let students access lower-order exercise representations.
Key results
- On ASSISTments at 70%/10%/20% split, DisenGCD reaches ACC 0.7334, RMSE 0.4209, AUC 0.7746, versus RCD at 0.7296, 0.4245, 0.7687.
- On Math at 70%/10%/20% split, DisenGCD reaches ACC 0.7582, RMSE 0.4004, AUC 0.7932, outperforming RCD (0.7534, 0.4034, 0.7866).
- Under 50% injected interaction noise on ASSISTments, the performance gap over RCD widens, per the paper's robustness experiments.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DisenGCD instead of RCD | Cognitive diagnosis · ASSISTments (60%/10%/30%) | AUC: 0.7663 -> 0.7723 | Helped |
| Use DisenGCD instead of RCD | Cognitive diagnosis · Math (70%/10%/20%) | AUC: 0.7866 -> 0.7932 | Helped |
| Learn all three representations only on the interaction graph rather than using three disentangled graphs (ablation) | DisenGCD · ASSISTments | AUC: 0.7723 -> 0.7678 | Hurt |
| Learn student representation on interaction graph and exercise/concept representations on relation graph rather than using three disentangled graphs (ablation) | DisenGCD · ASSISTments | AUC: 0.7723 -> 0.7701 | Hurt |
| Learn student and exercise representations on interaction graph and concept representation on relation graph (ablation) | DisenGCD · ASSISTments | AUC: 0.7723 -> 0.7678 | Hurt |
| Learn student and concept representations on interaction graph and exercise representation on relation graph (ablation) | DisenGCD · ASSISTments | AUC: 0.7723 -> 0.7685 | Hurt |
| Replace GAT with GCN (ablation) | DisenGCD · Math | AUC: 0.7932 -> 0.7639 | Hurt |
| Replace GAT with GraphSage (ablation) | DisenGCD · Math | AUC: 0.7932 -> 0.7881 | Hurt |
| Replace GAT with HAN (ablation) | DisenGCD · Math | AUC: 0.7932 -> 0.7647 | Hurt |
| Set meta multigraph hyper-node count P to 4 instead of 5 (ablation) | DisenGCD · ASSISTments | AUC: 0.7722 -> 0.7714 | Hurt |
| Set meta multigraph hyper-node count P to 6 instead of 5 (ablation) | DisenGCD · ASSISTments | AUC: 0.7722 -> 0.7711 | Hurt |
| Set meta multigraph hyper-node count P to 7 instead of 5 (ablation) | DisenGCD · ASSISTments | AUC: 0.7722 -> 0.7719 | Hurt |
Surprising findings
- Learned meta multigraphs transfer poorly: a meta multigraph learned on ASSISTments performs poorly on SLP, while those learned on Math transfer better to SLP.
Limitations
- Computational complexity grows with the number of hyper-nodes P and graph size, which may limit scalability to large datasets.
- Task transferability is limited to student-exercise-concept graphs.
Experimental setup
Datasets: ASSISTments Math SLP
Benchmarks: ASSISTments Math SLP
Metrics: ACC RMSE AUC
Baselines: DINA MIRT NCD ECD RCD SCD KSCD KaNCD
Builds on: RCD DiffMG GEMS PMMM GAT
Closest papers by meaning
- Collaborative Cognitive Diagnosis with Disentangled Representation Learning for Learner Modeling (NeurIPS 2024 · similarity 0.88)
- Disentangled Graph Self-supervised Learning for Out-of-Distribution Generalization (ICML 2024 · similarity 0.85)
- Graph-based Unsupervised Disentangled Representation Learning via Multimodal Large Language Models (NeurIPS 2024 · similarity 0.85)
- Symbolic Cognitive Diagnosis via Hybrid Optimization for Intelligent Education Systems (AAAI 2024 · similarity 0.84)
- Multi-Perspective Consolidation Enhanced Cognitive Diagnosis via Conditional Diffusion Model (AAAI 2025 · similarity 0.84)
- AD4CD: Causal-Guided Anomaly Detection for Enhancing Cognitive Diagnosis (AAAI 2025 · similarity 0.84)
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.