Research map / Probabilistic and scientific modeling / Molecules and proteins
AD4CD: Causal-Guided Anomaly Detection for Enhancing 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 AD4CD, a causal-inspired anomaly detection framework that uses response times and ECDF/skewness-based anomaly scores plus ELBO-based natural-ability reconstruction to model guessing and slipping in cognitive diagnosis models.
Problem
Existing cognitive diagnosis models treat all historical responses as normal, so anomalous student states and exercise properties (e.g., guessing, slipping, mislabeled concepts) bias diagnostic results.
Approach
Uses a causal graph and backdoor adjustment to motivate anomaly scores computed from response-time distributions (ECDF tails weighted by distribution skewness) for students and exercises, and a VAE-style reconstruction error as a natural-state anomaly score; these scores model guessing and slipping probabilities added on top of a general CDM backbone.
Key results
- Adding AD4CD to NCD improves ASSIST09 results from ACC 0.7262 / RMSE 0.4302 / AUC 0.7525 to ACC 0.7367 / RMSE 0.4216 / AUC 0.7672 (reported as significant, p<0.01).
- KANCD+AD4CD reaches ACC 0.7838 and AUC 0.7646 on Junyi, versus 0.7739 / 0.7373 for KANCD alone.
- Ablations on ASSIST09 show removing slipping modeling, guess modeling, or anomalous-state fusion each lowers ACC (0.7140, 0.7282, 0.7254 vs 0.7367).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add AD4CD to NCD | NCD · ASSIST09 | ACC: 0.7262 -> 0.7367 | Helped |
| Add AD4CD to NCD | NCD · ASSIST09 | RMSE: 0.4302 -> 0.4216 | Helped |
| Add AD4CD to NCD | NCD · ASSIST17 | AUC: 0.6923 -> 0.7195 | Helped |
| Add AD4CD to NCD | NCD · Junyi | RMSE: 0.4024 -> 0.3881 | Helped |
| Add AD4CD to cognitive diagnosis models | IRT, DINA, MIRT, NCD, KSCD, and KANCD · ASSIST09, ASSIST17, and Junyi | ACC, RMSE, and AUC | Helped |
| Add slipping modeling (ablation) | NCD+ · ASSIST09 | ACC: 0.7140 -> 0.7367 | Helped |
| Add slipping modeling (ablation) | NCD+ · ASSIST09 | RMSE: 0.4318 -> 0.4216 | Helped |
| Add slipping modeling (ablation) | NCD+ · ASSIST09 | AUC: 0.7524 -> 0.7672 | Helped |
| Add guessing modeling (ablation) | NCD+ · ASSIST09 | ACC: 0.7282 -> 0.7367 | Helped |
| Add guessing modeling (ablation) | NCD+ · ASSIST09 | RMSE: 0.4271 -> 0.4216 | Helped |
| Add guessing modeling (ablation) | NCD+ · ASSIST09 | AUC: 0.7577 -> 0.7672 | Helped |
| Add anomalous-state fusion modeling (ablation) | NCD+ · ASSIST09 | ACC: 0.7254 -> 0.7367 | Helped |
Surprising findings
- Performance gains in ACC are relatively smaller for some models while AUC improves more, which the authors attribute to data distribution imbalance.
Limitations
- Anomalies are not directly observable, so validation relies on indirect evidence such as robustness tests with simulated noise and visualizations.
- Gains in ACC are modest for some backbones (e.g., small improvements on ASSIST17 for some models).
Experimental setup
Datasets: ASSIST09 ASSIST17 Junyi
Benchmarks: ASSIST09 ASSIST17 Junyi
Metrics: ACC RMSE AUC
Baselines: IRT MIRT DINA NCD KSCD KANCD
Builds on: NCD KSCD KANCD DINA IRT
Closest papers by meaning
- Towards Accurate and Fair Cognitive Diagnosis via Monotonic Data Augmentation (NeurIPS 2024 · similarity 0.88)
- Zero-1-to-3: Domain-Level Zero-Shot Cognitive Diagnosis via One Batch of Early-Bird Students towards Three Diagnostic Objectives (AAAI 2024 · similarity 0.88)
- Knowledge Is Power: Harnessing Large Language Models for Enhanced Cognitive Diagnosis (AAAI 2025 · similarity 0.87)
- Symbolic Cognitive Diagnosis via Hybrid Optimization for Intelligent Education Systems (AAAI 2024 · similarity 0.87)
- Boosting Neural Cognitive Diagnosis with Student’s Affective State Modeling (AAAI 2024 · similarity 0.87)
- Enhancing Cognitive Diagnosis Using Un-interacted Exercises: A Collaboration-Aware Mixed Sampling Approach (AAAI 2024 · 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.