Research map / Probabilistic and scientific modeling / Molecules and proteins
Symbolic Cognitive Diagnosis via Hybrid Optimization for Intelligent Education Systems
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 a symbolic cognitive diagnosis (SCD) framework that alternates genetic-programming symbolic regression of a monotonic interaction function with Adam optimization of student and exercise parameters, improving interpretability while keeping generalization competitive on four education datasets.
Problem
Cognitive diagnosis models face a trade-off between generalization (predicting student responses) and interpretability (interaction function and proficiency levels); existing models either lack interpretable non-linear interaction functions or are black-box neural networks.
Approach
Represent the student-exercise interaction function as a symbolic tree learned by genetic programming with a monotonic operator set, and alternate this with gradient-based Adam optimization of proficiency and exercise parameters in a hybrid discrete-continuous loop.
Key results
- SCDM achieves the best F1 on all four datasets (e.g., 74.14 on Math1 vs. 73.27 for KaNCD) and the best DOA on all four (e.g., 81.78 on FracSub vs. 80.72 for NCDM).
- SCDM's accuracy and AUC are competitive but not uniformly best; KaNCD and IRT lead on some accuracy/AUC columns (e.g., KaNCD 68.23 accuracy on Math1 vs. SCDM 67.78).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add GP to learn and update the interaction function (ablation) | SCDM · Math1 | accuracy: 67.22 -> 67.78 | Helped |
| add Adam instead of using an evolutionary strategy for parameter optimization (ablation) | SCDM · Math1 | accuracy: 67.30 -> 67.78 | Helped |
| add GP to learn and update the interaction function (ablation) | SCDM · Math2 | accuracy: 69.05 -> 69.79 | Helped |
| add Adam instead of using an evolutionary strategy for parameter optimization (ablation) | SCDM · Math2 | accuracy: 68.43 -> 69.79 | Helped |
| add GP to learn and update the interaction function (ablation) | SCDM · FracSub | accuracy: 82.67 -> 83.26 | Helped |
| add Adam instead of using an evolutionary strategy for parameter optimization (ablation) | SCDM · FracSub | accuracy: 76.42 -> 83.26 | Helped |
| add GP to learn and update the interaction function (ablation) | SCDM · NeurIPS2020 | accuracy: 63.43 -> 67.25 | Helped |
| add Adam instead of using an evolutionary strategy for parameter optimization (ablation) | SCDM · NeurIPS2020 | accuracy: 63.27 -> 67.25 | Helped |
Surprising findings
- Ablation shows removing GP (using only the manual MIRT-like initial function) or replacing Adam with evolutionary strategy both reduce performance, indicating each component matters.
- DINA attains a high DOA but very poor generalization, illustrating the interpretability-generalization trade-off.
Limitations
- Interpretability of symbolic trees is not quantified beyond DOA and a single qualitative case study.
- Accuracy and AUC are not uniformly best; generalization is only competitive, not superior.
- Hyperparameters such as GP population size and generations are fixed by hand and the framework's convergence is not theoretically analyzed.
Experimental setup
Datasets: Math1 Math2 FracSub NeurIPS2020
Benchmarks: Math1 Math2 FracSub NeurIPS2020
Metrics: accuracy AUC F1 DOA
Baselines: IRT MIRT DINA NCDM KaNCD SCDM w.o. GP SCDM w.o. Adam
Builds on: item response theory multidimensional IRT DINA NCDM KaNCD
Closest papers by meaning
- AD4CD: Causal-Guided Anomaly Detection for Enhancing 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)
- Towards Accurate and Fair Cognitive Diagnosis via Monotonic Data Augmentation (NeurIPS 2024 · similarity 0.86)
- Boosting Neural Cognitive Diagnosis with Student’s Affective State Modeling (AAAI 2024 · similarity 0.85)
- Enhancing Cognitive Diagnosis Using Un-interacted Exercises: A Collaboration-Aware Mixed Sampling Approach (AAAI 2024 · similarity 0.85)
- DisenGCD: A Meta Multigraph-assisted Disentangled Graph Learning Framework for Cognitive Diagnosis (NeurIPS 2024 · 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.