Research map / Label-efficient and robust learning / Recognition and long-tail
Learning Concept Prerequisite Relation via Global Knowledge Relation Optimization
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 GKROM, a heterogeneous graph attention network that jointly learns concept prerequisite, document-concept containment, and document-document dependency relations via multi-task loss to improve concept prerequisite prediction.
Problem
Predicting concept prerequisite relations in educational knowledge networks, where other knowledge relation types (document-concept containment, document-document dependency) are typically ignored even though they influence each other.
Approach
Treat concept prerequisite prediction as the main task within a multi-objective framework over a document-concept heterogeneous graph, with auxiliary containment and dependency prediction tasks that share GAT-encoded node embeddings and are balanced by a weighted global loss.
Key results
- On LectureBank GKROM reaches ACC 0.8232, about 1.22% above the best baseline HGAPNet (0.8110).
- On University Course GKROM achieves F1 0.8740, about 0.3% above HGAPNet, but its AUC (0.9395) is slightly below HGAPNet (0.9456).
- Ablation on LectureBank shows removing both auxiliary tasks drops ACC from 0.8232 to 0.8049 and AUC from 0.9054 to 0.8978.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use GKROM instead of HGAPNet | concept prerequisite prediction · University Course | ACC: 0.8687 -> 0.8698 | Helped |
| use GKROM instead of HGAPNet | concept prerequisite prediction · University Course | F1: 0.8710 -> 0.8740 | Helped |
| use GKROM instead of HGAPNet | concept prerequisite prediction · University Course | AUC: 0.9456 -> 0.9395 | Hurt |
| use GKROM instead of HGAPNet | concept prerequisite prediction · LectureBank | ACC: 0.8110 -> 0.8232 | Helped |
| use GKROM instead of HGAPNet | concept prerequisite prediction · LectureBank | F1: 0.8098 -> 0.8199 | Helped |
| use GKROM instead of HGAPNet | concept prerequisite prediction · LectureBank | AUC: 0.9017 -> 0.9054 | Helped |
| use GKROM instead of HGAPNet | concept prerequisite prediction · MOOC | ACC: 0.8750 -> 0.8796 | Helped |
| use GKROM instead of HGAPNet | concept prerequisite prediction · MOOC | F1: 0.8787 -> 0.8847 | Helped |
| use GKROM instead of HGAPNet | concept prerequisite prediction · MOOC | AUC: 0.9388 -> 0.9417 | Helped |
| add document-concept containment task (ablation) | GKROM · LectureBank | ACC: 0.8171 -> 0.8232 | Helped |
| add document-document dependency task (ablation) | GKROM · LectureBank | ACC: 0.8110 -> 0.8232 | Helped |
| add document-concept containment and document-document dependency tasks (ablation) | GKROM · LectureBank | ACC: 0.8049 -> 0.8232 | Helped |
Surprising findings
- GKROM's AUC on University Course is slightly lower than HGAPNet's, which the authors attribute to optimizing overall performance rather than one metric.
- Improvements over the strongest baseline are small (mostly under 1.3%), with the University Course ACC gain only 0.11%.
Limitations
- Improvements over baselines are marginal on several metrics and AUC is lower than HGAPNet on University Course.
- Edges in the heterogeneous graph rely on hand-set similarity thresholds (PMI, TF-IDF, cosine) and GloVe embeddings rather than learned representations.
- Only three small public datasets evaluated, with negatives randomly sampled.
Experimental setup
Datasets: University Course LectureBank MOOC
Benchmarks: University Course LectureBank MOOC
Metrics: ACC F1 AUC
Baselines: RefD M3 GAE VGAE PREREQ R-VGAE(T) R-VGAE(P) MHAVGAE HGAPNet NB SVM RF
Builds on: VGAE R-VGAE MHAVGAE HGAPNet graph attention network
Closest papers by meaning
- Toward a Unified Statistical Theory of Unsupervised Pretraining and Supervised Neural Knowledge Graph Learning (NeurIPS 2026 · similarity 0.85)
- Towards Global-Topology Relation Graph for Inductive Knowledge Graph Completion (AAAI 2025 · similarity 0.83)
- DocKS-RAG: Optimizing Document-Level Relation Extraction through LLM-Enhanced Hybrid Prompt Tuning (ICML 2025 · similarity 0.83)
- Reasoning of Large Language Models over Knowledge Graphs with Super-Relations (ICLR 2025 · similarity 0.83)
- Structure Is All You Need: Structural Representation Learning on Hyper-Relational Knowledge Graphs (ICML 2025 · similarity 0.83)
- Revisiting Document-Level Relation Extraction with Context-Guided Link Prediction (AAAI 2024 · 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.