Research map / Image generation / Fast diffusion sampling
Multi-Perspective Consolidation Enhanced Cognitive Diagnosis via Conditional Diffusion Model
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 DMC-CDM, a cognitive diagnosis model that adds a semantic extractor to collaborative features and consolidates multi-perspective learner states via a conditional diffusion model motivated by mutual-information maximization.
Problem
Cognitive diagnosis from learner interaction logs is ill-posed due to information loss from under-expressive interaction functions and incomplete (partially observed) interactions.
Approach
Extract cognitive states from a single perspective using collaborative plus semantic (BGE-M3) embeddings fused by cross-attention, then retrieve similar learners' logs as other perspectives and use a conditional diffusion transformer to reconstruct the observed learner's states conditioned on them, maximizing mutual information between perspectives; predictions are a cosine-similarity-weighted ave
Key results
- DMC-CDM reaches AUC 0.8834 on Junyi and 0.8500 on Ifly, versus best baselines of 0.8671 (CMES) and 0.8436 (CMES) respectively.
- On MoocRadar, DMC-CDM attains the best ACC (0.9143) and F1 (0.9521), while RCD has the highest AUC (0.9241).
- Ablation shows removing multi-perspective consolidation hurts Ifly AUC more (0.7679) than removing the semantic extractor (0.8097).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DMC-CDM instead of RCD | Cognitive diagnosis · MoocRadar | ACC: 0.9126 -> 0.9143* | Helped |
| Use DMC-CDM instead of NeuralCDM+ | Cognitive diagnosis · Junyi | AUC: 0.8826 -> 0.8834 | Helped |
| Use DMC-CDM instead of CMES | Cognitive diagnosis · Ifly | F1: 0.7989 -> 0.8079* | Helped |
| Add multi-perspective consolidation (ablation) | DMC-CDM · Junyi | AUC: 0.8574 -> 0.8834 | Helped |
| Add multi-perspective consolidation (ablation) | DMC-CDM · Ifly | AUC: 0.7679 -> 0.8500* | Helped |
| Add multi-perspective consolidation (ablation) | DMC-CDM · MoocRadar | AUC: 0.9069 -> 0.9064 | No clear effect |
| Add semantic extractor (ablation) | DMC-CDM · Junyi | AUC: 0.8626 -> 0.8834 | Helped |
| Add semantic extractor (ablation) | DMC-CDM · Ifly | AUC: 0.8097 -> 0.8500* | Helped |
| Add semantic extractor (ablation) | DMC-CDM · MoocRadar | AUC: 0.9010 -> 0.9064 | Helped |
| Add multi-perspective consolidation and semantic extractor (ablation) | DMC-CDM · Junyi | AUC: 0.8291 -> 0.8834 | Helped |
| Add multi-perspective consolidation and semantic extractor (ablation) | DMC-CDM · Ifly | AUC: 0.7146 -> 0.8500* | Helped |
| Add multi-perspective consolidation and semantic extractor (ablation) | DMC-CDM · MoocRadar | AUC: 0.9076 -> 0.9064 | No clear effect |
Surprising findings
- On MoocRadar, RCD (AUC 0.9241) and the proposed model's AUC (0.9064) are not the top AUC, so the method does not win on every metric.
- Multi-perspective consolidation contributed more than the semantic extractor in ablations.
Limitations
- Theoretical justification relies on Assumption 1 (latent states recoverable from all possible observations) and a loose mutual-information bound.
- Performance gains are not uniform across datasets and metrics; RCD has higher MoocRadar AUC.
- Retrieval of similar learners as other perspectives may be less informative in sparse data, and the paper notes diminishing returns for large K.
Experimental setup
Datasets: MoocRadar Ifly Junyi
Benchmarks: MoocRadar Ifly Junyi
Metrics: AUC ACC F1
Baselines: IRT MIRT NeuralCDM NeuralCDM+ RCD KaNCD TechCD EIRS CMES
Builds on: NeuralCDM Diffusion Transformer (DiT) guided diffusion for sequential recommendation (Yang et al. 2024) DDPM
Closest papers by meaning
- Knowledge Is Power: Harnessing Large Language Models for Enhanced Cognitive Diagnosis (AAAI 2025 · similarity 0.92)
- Collaborative Cognitive Diagnosis with Disentangled Representation Learning for Learner Modeling (NeurIPS 2024 · similarity 0.90)
- Enhancing Cognitive Diagnosis Using Un-interacted Exercises: A Collaboration-Aware Mixed Sampling Approach (AAAI 2024 · similarity 0.89)
- 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)
- Boosting Neural Cognitive Diagnosis with Student’s Affective State Modeling (AAAI 2024 · similarity 0.87)
- Towards Accurate and Fair Cognitive Diagnosis via Monotonic Data Augmentation (NeurIPS 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.