Aspect-Aware Content-Based Recommendations for Mathematical Research Papers

Content-based research paper recommendation (CbRPR) has been limited to computer science and biomedicine, where textual or citation similarity works well. Mathematics papers, however, connect through shared proof techniques, logical implications, or natural generalizations—links that existing methods miss. To close this gap, a team led by Ankit Satpute and Bela Gipp first conducted an expert study revealing that mathematical relevance is inherently aspect-driven. They then built GoldRiM (expert-annotated) and SilverRiM (automatically derived), the first datasets for aspect-aware CbRPR in mathematics.

Recognizing that plain LLM embeddings fail to capture mathematical content, the team designed AchGNN, an aspect-conditioned heterogeneous graph neural network. Unlike standard methods, AchGNN jointly models three signals: textual semantics, citation structure, and author lineage. Across both GoldRiM and SilverRiM datasets, AchGNN consistently outperformed prior aspect-based approaches, achieving substantial gains across all evaluated aspects. Ablation studies confirmed the contributions of individual supervision signals, authorship lineage, and graph structure.

The researchers further validated AchGNN on the Papers with Code dataset (machine learning publications), demonstrating that the aspect-aware approach transfers effectively beyond mathematics. The system is now deployed on the MaRDI platform, giving mathematicians a practical tool to discover conceptually related papers rather than just textually similar ones. All datasets and code are released publicly for reproducibility. The work is accepted at SIGIR 2026.