Research map / Probabilistic and scientific modeling / Molecules and proteins
Predictive, scalable and interpretable knowledge tracing on structured domains
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 PSI-KT, a hierarchical probabilistic state-space model of knowledge tracing with latent learner traits, an Ornstein-Uhlenbeck knowledge-state process coupled to a learned prerequisite graph, and variational continual learning for scalable, interpretable prediction.
Problem
Knowledge tracing for intelligent tutoring systems requires predictive accuracy, scalability to new interactions, and interpretable learner traits and prerequisite structure, which existing deep and psychological models trade off.
Approach
A three-level hierarchical state-space model where learner-specific cognitive traits drive OU-process knowledge states coupled through a shared, factorized prerequisite graph, with amortized variational inference and variational continual learning so new interactions update the posterior without retraining.
Key results
- PSI-KT achieves within-learner prediction of .68 on Assist12 and .83 on Junyi15, the best among all compared models.
- PSI-KT attains the best between-learner generalization on Assist12 (.61), Assist17 (.53) and Junyi15 (.79), and wins with fine-tuning on Assist17 (.56).
- On Junyi15 prerequisite alignment, PSI-KT reaches MRR .0086 and crowd Jaccard .0095 vs best baseline .0082 and .0047, and its causal-support regression coefficient is 0.97 (p<.001) vs 0.42 (p=.593) for the best baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PSI-KT instead of AKT | Knowledge tracing · Assist12 within-learner | accuracy: .67 -> .68 | Helped |
| Use PSI-KT instead of GKT | Knowledge tracing · Assist12 between-learner | accuracy: .61 -> .61 | No clear effect |
| Use PSI-KT instead of GKT | Knowledge tracing · Assist12 between-learner with fine-tuning | accuracy: .62 -> .62 | No clear effect |
| Use PSI-KT instead of QIKT | Knowledge tracing · Assist17 within-learner | accuracy: .58 -> .63 | Helped |
| Use PSI-KT instead of GKT | Knowledge tracing · Assist17 between-learner | accuracy: .53 -> .53 | No clear effect |
| Use PSI-KT instead of GKT | Knowledge tracing · Assist17 between-learner with fine-tuning | accuracy: .54 -> .56 | Helped |
| Use PSI-KT instead of AKT | Knowledge tracing · Junyi15 within-learner | accuracy: .81 -> .83 | Helped |
| Use PSI-KT instead of GKT | Knowledge tracing · Junyi15 between-learner | accuracy: .77 -> .79 | Helped |
| Use PSI-KT instead of DKTF | Knowledge tracing · Junyi15 between-learner with fine-tuning | accuracy: .84 -> .80 | Hurt |
| Use PSI-KT inferred prerequisite graph instead of the best baseline graph | Prerequisite graph inference · Junyi15 | Jaccard similarity, crowd: .0047 -> .0095 | Helped |
| Use PSI-KT inferred prerequisite graph instead of the best baseline graph | Prerequisite graph inference · Junyi15 | nLL: 3.03 -> 4.11 | Hurt |
| Use PSI-KT inferred edge probabilities instead of the best baseline predictions | Causal-support prediction · Assist12 | regression coefficient: 1.05 -> 1.15 | Helped |
Surprising findings
- Baselines needed training data from around 60k learners to match PSI-KT's accuracy trained on small cohorts, per the discussion.
- On Junyi15 with fine-tuning, DKTF (.84) slightly outperforms PSI-KT (.80) in between-learner generalization.
Limitations
- Evaluation limited to pre-college mathematics datasets.
- Uses exponential forgetting from an OU process, so alternative memory decay forms are not explored.
- Prerequisite graph is not constrained to be acyclic or otherwise structurally regularized.
Experimental setup
Datasets: ASSISTments 2012 (Assist12) ASSISTments 2017 (Assist17) Junyi Academy 2015 (Junyi15)
Benchmarks: Assist12 Assist17 Junyi15
Metrics: AUC (within- and between-learner prediction) mutual information specificity consistency disentanglement KL MRR Jaccard similarity negative log-likelihood regression coefficients for causal support
Baselines: HLR PPE DKT DKTF HKT AKT GKT QIKT
Builds on: Hawkes knowledge tracing half-life regression deep knowledge tracing variational continual learning Bayesian causal induction
Closest papers by meaning
- Uncertainty-aware Knowledge Tracing (AAAI 2025 · similarity 0.87)
- Cooperative Knowledge Distillation: A Learner Agnostic Approach (AAAI 2024 · similarity 0.86)
- Why Knowledge Distillation Works in Generative Models: A Minimal Working Explanation (NeurIPS 2025 · similarity 0.85)
- K12-KGraph: A Curriculum-Aligned Knowledge Graph for Benchmarking and Training Educational LLMs (NeurIPS 2026 · similarity 0.85)
- Speculative Knowledge Distillation: Bridging the Teacher-Student Gap Through Interleaved Sampling (ICLR 2025 · similarity 0.85)
- Can Students Beyond the Teacher? Distilling Knowledge from Teacher’s Bias (AAAI 2025 · similarity 0.85)
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.