Research map / Label-efficient and robust learning / Recognition and long-tail
Neighborhood-Aware Negative Sampling for Student Knowledge and Behavior Modeling
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 Neighborhood-Aware Negative Sampling (NANS) built on a weighted question transition graph, plus KoBeM, a multi-task memory-augmented model jointly predicting student next-question choice and performance on EdNet and Junyi.
Problem
Uniform random negative sampling is uninformative for next-question prediction in student sequences with a very large question space, and student knowledge and behavior are modeled separately despite their interrelation.
Approach
Build a weighted question transition graph from student sequences, take each question's top-N highest-weight neighbors as its candidate negative set, and sample negatives uniformly from that set; pair this with a dual-MANN knowledge/behavior model trained with Pareto multi-task learning.
Key results
- On EdNet, NANS-KoBeM reaches HR 0.0982 / NDCG 0.0622 / MRR 0.0503 for next-question prediction, slightly above GNNO-KoBeM (0.0974 HR).
- On Junyi, NANS-KoBeM achieves HR 0.4919 vs 0.4875 for the best negative-sampling baseline (GNNO-KoBeM).
- For student performance prediction, NANS-KoBeM achieves AUC 0.6615 on EdNet and 0.8779 on Junyi, above DKT (0.6393 and 0.8623).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use NANS-KoBeM instead of GNNO-KoBeM | KoBeM · EdNet | HR: 0.0974 -> 0.0982 | Helped |
| Use NANS-KoBeM instead of GNNO-KoBeM | KoBeM · EdNet | NDCG: 0.0618 -> 0.0622 | Helped |
| Use NANS-KoBeM instead of GNNO-KoBeM | KoBeM · Junyi | HR: 0.4875 -> 0.4919 | Helped |
| Use NANS-KoBeM instead of GNNO-KoBeM | KoBeM · Junyi | MRR: 0.3389 -> 0.3422 | Helped |
| Add knowledge modeling (ablation) | NANS-KoBeM · EdNet | HR: 0.0877 -> 0.0982 | Helped |
| Add knowledge modeling (ablation) | NANS-KoBeM · Junyi | HR: 0.4391 -> 0.4919 | Helped |
| Add behavior modeling (ablation) | NANS-KoBeM · EdNet | AUC: 0.6296 -> 0.6615 | Helped |
| Add behavior modeling (ablation) | NANS-KoBeM · Junyi | AUC: 0.8558 -> 0.8779 | Helped |
Surprising findings
- Negatives drawn from the highest-weight neighbors (MCNS/GNNO-style or weighted sampling) can perform similarly to or worse than random sampling, as the authors note that high-weight neighbors reduce diversity (WD-KoBeM worse on Junyi).
Limitations
- Gains over the strongest negative-sampling baselines (GNNO-KoBeM, MCNS-KoBeM) are small, often within ~0.01 absolute and borderline in significance.
- Evaluated only on two education datasets; generalization to other sequential decision tasks is claimed but not tested.
- Next-question evaluation uses 99 randomly sampled negatives rather than the full question set.
Experimental setup
Datasets: EdNet Junyi
Benchmarks: EdNet Junyi
Metrics: HR@5 NDCG@5 MRR@5 AUC
Baselines: LSTM MANN RD-KoBeM WD-KoBeM MCNS-KoBeM GNNO-KoBeM DKT DeepIRT SAKT SAINT AKT DKVMN
Builds on: DKVMN MCNS GNNO Pareto MTL
Closest papers by meaning
- Simplify and Robustify Negative Sampling for Implicit Collaborative Filtering (NeurIPS 2020 · similarity 0.87)
- Adaptive Hardness Negative Sampling for Collaborative Filtering (AAAI 2024 · similarity 0.85)
- Enhancing Student Performance Prediction on Learnersourced Questions with SGNN-LLM Synergy (AAAI 2024 · similarity 0.83)
- PLD: A Choice-Theoretic List-Wise Knowledge Distillation (NeurIPS 2025 · similarity 0.83)
- Uncertainty-aware Knowledge Tracing (AAAI 2025 · similarity 0.83)
- Personalized Exercise Recommendation with Semantically-Grounded Knowledge Tracing (NeurIPS 2025 · similarity 0.83)
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.