"Aspect-Aware Content-Based Recommendations for Mathematical Research Papers"

Ankit Satpute* , Andre Greiner-Petter , Noah Gießing , Olaf Teschke, Moritz Schubotz, Akiko Aizawa and Bela Gipp.

*Corresponding Author

🚀 Accepted to the Full Papers Track of SIGIR 2026

Abstract

This repository contains an aspect-aware content-based research paper recommendation system code and dataset designed specifically for mathematics, where relatedness is often conceptual rather than based on textual similarity or citation overlap. Unlike existing approaches that work well in domains such as computer science or biomedicine, this project addresses the unique challenges of mathematical literature by modeling connections through shared proof techniques, logical implications, and theoretical generalizations. The project introduces GoldRiM and SilverRiM, the first datasets for aspect-aware mathematical paper recommendation, and presents AchGNN, an aspect-conditioned heterogeneous graph neural network that integrates textual semantics, citation networks, and author relationships. Experimental results show that AchGNN consistently outperforms prior recommendation methods across multiple aspects and also generalizes effectively to machine learning literature. The system has been deployed on the MaRDI platform An example document with recommendations to support mathematical research discovery.

Repository Structure

Path	Content
data/	GoldRiM and SilverRiM: description, download and loading
zbmath_hetnet/	Harvesting zbMATH Open and building the author–document citation network
achgnn/	AchGNN: graph construction, node features, training, evaluation and recommendation
slurm/	Job scripts for running the full pipeline on a GPU cluster
src/	Baseline indexers, data analysis and ablation scripts
material/	Supplementary material: related datasets, aspect definitions, additional plots

Installation

Python 3.10 or newer is required. Install PyTorch for your platform first (pytorch.org; CUDA builds for GPU training), then the pipeline dependencies:

python -m venv .venv
source .venv/bin/activate            # Windows: .venv\Scripts\activate
pip install torch                    # or a CUDA build, e.g. --index-url https://download.pytorch.org/whl/cu124
pip install -r achgnn/requirements.txt -r zbmath_hetnet/requirements.txt

The scripts in src/ use the pinned environment in requirements.txt.

Data

The repository provides two datasets:

• SilverRiM: automatically derived aspect-aware recommendations (training and evaluation)
• GoldRiM: expert-curated recommendations (evaluation)

Both are distributed as a single CSV file of zbMATH Open document pairs. See data/Readme.md for the download link, the file format and examples. The pipeline below expects the file at data/mathRec_dataset.csv.

Pipeline

AchGNN operates on a heterogeneous graph of zbMATH Open papers and authors. The complete pipeline consists of five steps:

Step	Command	Output
1. Harvest zbMATH Open	python zbmath_hetnet/crawl.py all --out-dir experiments/zbmath	raw API records
2. Build the networks	python achgnn/dataset.py, python zbmath_hetnet/build_graph.py --with-text, python achgnn/prepare_graph.py	paper–author graph, labelled pairs
3. Node features	python achgnn/embed.py --encoder hf --devices all	Qwen3 embeddings of papers and authors
4. Training	torchrun --nproc_per_node=<GPUs> achgnn/train.py	trained model, metrics
5. Evaluation and recommendation	achgnn/evaluate.py, achgnn/summarize.py, achgnn/recommend.py	P@10, R@10, MRR; aspect-specific recommendations

Commands with all arguments are given in achgnn/README.md. Every step runs on CPU or GPU; steps 3 and 4 use all GPUs of a node when available.

A small-scale run (a subset of SilverRiM, all GoldRiM pairs and the zbMATH documents they involve, CPU only) executes the complete pipeline end to end in a few minutes:

bash achgnn/run_sample.sh data/mathRec_dataset.csv

Running on a SLURM Cluster

The slurm/ directory contains one job script per pipeline step, configured for a node with 4 × NVIDIA A100 80 GB.

1. Adjust partition, account and paths in slurm/config.sh.
2. On a login node with internet access, create the environment and download the dataset and the Qwen3 encoder weights:

   bash slurm/setup_env.sh

3. Submit the pipeline from the repository root as a chain of dependent jobs:

   bash slurm/submit_all.sh                # crawl → prepare → embed → train (5 folds) → evaluate
   bash slurm/submit_all.sh --skip-crawl   # when the zbMATH crawl is already available

Script	Resources	Step
01_crawl.sbatch	CPU, internet access	harvest zbMATH Open (resumable)
02_prepare.sbatch	CPU, 192 GB RAM	dataset pairs, zbMATH network, AchGNN graph
03_embed.sbatch	4 GPUs	Qwen3-Embedding-8B features, one worker per GPU (resumable)
04_train.sbatch	4 GPUs, job array 0–4	DDP training and evaluation, one task per cross-validation fold (resumable)
05_evaluate.sbatch	4 GPUs	evaluation of all folds and mean ± std summary

Individual steps can be submitted on their own, e.g. sbatch --partition=<partition> --array=0 slurm/04_train.sbatch. Logs are written to slurm/logs/, results to experiments/work/runs/.

Supplementary Materials

We offer supplementary materials in (material), such as a summary table of existing CbRPR datasets, alternative visualizations of the results, definitions of aspect labels, and additional plots.

Citation

If you're using or referring to our paper in your research or applications, please cite using this BibTeX:

@inproceedings{Satpute2026,
	title     = {Aspect-Aware Content-Based Recommendations for Mathematical Research Papers},
	author    = {Satpute, Ankit and Greiner-Petter, Andre and Giessing, Noah and Teschke, Olaf and Schubotz, Moritz and Aizawa, Akiko and Gipp, Bela},
	year      = 2026,
	month     = {July.},
	booktitle = {Proceedings of 49th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR ’26)},
	publisher = {ACM},
	address   = {Melbourne | Naarm, Australia},
	topic     = {rec}
}

License

CC-BY-SA 4.0. The dataset includes non-copyrighted bibliographic metadata and reference data derived from I4OSC (CC0).