Research map / Graph learning / Knowledge and temporal graphs
Multi-type MOOCs Recommendation: Leveraging Deep Multi-Relational Representation and Hierarchical Reasoning
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 TOME, a two-stage multi-type MOOC recommender that fuses course, knowledge-concept and dynamic video graphs via GNNs and a multi-relation adaptive LSTM, then recommends courses, concepts and videos with hierarchical beam search.
Problem
Recommending multiple types of MOOC content (courses, knowledge concepts, videos) jointly, while capturing the interrelations among these content types and the hierarchical learner-content interactions.
Approach
Builds course-relational and knowledge-concept-relational static graphs plus a dynamic video-relational graph from browsing sequences, encodes each with GNNs (DGAT/GCN), fuses them with a multi-relation adaptive LSTM, and performs top-down hierarchical beam search over joint log-probabilities across the three levels, with an optional personalization/diversification layer for video recommendations.
Key results
- On Computer, TOME reaches HR@5 0.9484 for course, 0.9427 for video and 0.7959 for concepts... (table reports HR@5 0.9484 course-level, 0.9427 video-level on Computer), outperforming all single-type and multi-task baselines.
- On MOOC3, TOME achieves HR@5 0.7707 for course-level and 0.7959 for knowledge concept-level recommendation (Table 2 values), above the best baseline MCLRec (0.6717 and 0.7585).
- Ablations show that integrating multiple relational graphs generally beats single-graph variants, and beam width beta=100 performs best on most levels.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TOME instead of MCLRec | course recommendation · Computer | HR@5: 0.9336 -> 0.9484 | Helped |
| Use TOME instead of MCLRec | course recommendation · MOOC3 | HR@5: 0.9306 -> 0.9427 | Helped |
| Use TOME instead of MCLRec | video recommendation · Computer | HR@5: 0.6717 -> 0.7707 | Helped |
| Use TOME instead of MCLRec | video recommendation · MOOC3 | HR@5: 0.7585 -> 0.7959 | Helped |
| Use TOME instead of CoSeRec | knowledge concept recommendation · Computer | FACC: 0.2678 -> 0.6600 | Helped |
| Use TOME instead of SRMA | knowledge concept recommendation · MOOC3 | FACC: 0.0295 -> 0.4703 | Helped |
| Remove the course-relational graph representation (ablation) | TOME video recommendation · Computer and MOOC3 | recommendation performance | Hurt |
| Remove the knowledge concept-relational graph representation (ablation) | TOME video recommendation · Computer and MOOC3 | recommendation performance | Hurt |
| Remove the video-relational graph representation (ablation) | TOME video recommendation · Computer and MOOC3 | recommendation performance | Hurt |
| Use only course and knowledge concept graph representations (ablation) | TOME video recommendation · Computer and MOOC3 | recommendation performance | Hurt |
| Increase hierarchical beam width to β=100 (ablation) | TOME hierarchical beam search · Most dataset and recommendation-level settings | recommendation performance | Helped |
| Set hierarchical beam width to β=10 instead of β=100 (ablation) | TOME knowledge concept recommendation · Computer | recommendation performance | No clear effect |
Surprising findings
- Multi-task baselines (MMoE, SNR, PLE) do not stand out and sometimes underperform single-type sequential models.
- Knowledge-concept recommendation performance is much higher on Computer than on MOOC3 for all models.
Limitations
- Relies on static domain-knowledge graphs and sequential browsing data, limiting generality to other platforms.
- Multi-task baselines were adapted using TOME representations, so comparison is indirect.
- Improvements over course-level recommendation are less pronounced where the number of courses is small.
Experimental setup
Datasets: Computer (Xuetang X) MOOC3 (Xuetang X)
Benchmarks: Computer MOOC3
Metrics: HR@5 NDCG@5 MRR FACC N2 diversity metrics
Baselines: COTREC lightGCN SR-GNN CoSeRec SRMA S3Rec TiCoSeRec IntGNN MCLRec MMoE SNR PLE
Builds on: Dimensional Graph Attention Network GCN LSTM beam search
Closest papers by meaning
- MRMLREC: A Two-Stage Approach for Addressing Data Sparsity in MOOC Video Recommendation (Student Abstract) (AAAI 2024 · similarity 0.91)
- Personalized Exercise Recommendation with Semantically-Grounded Knowledge Tracing (NeurIPS 2025 · similarity 0.84)
- EduLLM: Leveraging Large Language Models and Framelet-Based Signed Hypergraph Neural Networks for Student Performance Prediction (ICML 2025 · similarity 0.82)
- GenAL: Generative Agent for Adaptive Learning (AAAI 2025 · similarity 0.82)
- K12-KGraph: A Curriculum-Aligned Knowledge Graph for Benchmarking and Training Educational LLMs (NeurIPS 2026 · similarity 0.82)
- STEM: Unleashing the Power of Embeddings for Multi-Task Recommendation (AAAI 2024 · similarity 0.82)
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.