Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
SignFlow Bipartite Subgraph Network For Large-Scale Graph Link Sign Prediction
Authors: Yixiao Zhou, Xiaoqing Lyu, Hongxiang Lin, Huiying Hu, Tuo Wang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments reveal that SBSN shows superior performance in both mediumand large-scale datasets, efficiently managing memory and computational resources, making it a scalable solution for extensive applications. 5 Experiments In this section, we introduce seven real-world datasets and baseline models, along with our experimental setup. We then present the average metrics, demonstrating the performance of our model, after which we discuss the results in memory management and conduct ablation studies and parameter studies to evaluate the scalability of our model on large-scale datasets. |
| Researcher Affiliation | Academia | Yixiao Zhou Wangxuan Institute of Computer Technology Peking University Hongxiang Lin Wangxuan Institute of Computer Technology Peking University Huiying Hu Wangxuan Institute of Computer Technology Peking University Tuo Wang Wangxuan Institute of Computer Technology Peking University Xiaoqing Lyu Wangxuan Institute of Computer Technology, State Key Laboratory of Multimedia Information Processing Peking University Correspondence: EMAIL. Author emails: EMAIL, EMAIL, EMAIL, EMAIL, EMAIL |
| Pseudocode | No | The paper describes its methods in prose within sections like '4 Proposed Method', '4.1 Topological Subgraph Extractor', '4.2 Directed Sign Flow Passing', and '4.3 Sign Flow Aggregator', using mathematical formulations and descriptive text, but no explicit pseudocode blocks or algorithms are presented. |
| Open Source Code | Yes | The implementation of SBSN is publicly available at https://github.com/WICTSA/SBSN. |
| Open Datasets | Yes | We conduct a series of experiments on four small-scale and medium-scale real-world datasets: Bonanza2, U.S. House[5], ML-1M3 and Amazon-Book4, and three large-scale datasets: ML-10M2, ML-32M2 and Amazon-Book-51M5. 2https://www.bonanza.com/ 3https://grouplens.org/datasets/movielens/ 4https://jmcauley.ucsd.edu/data/amazon/index.html 5https://jmcauley.ucsd.edu/data/amazon_v2/index.html |
| Dataset Splits | Yes | During the experiments, we follow the experimental setup in [5], randomly splitting the links of each dataset into three parts: 10% for testing, 5% for validation, and the remaining 85% for training. For the large-scale dataset, we split the data into 8% for testing, 2% for validation, and 90% for training, and average the results over three runs. To ensure the stability and reliability of the experimental results, We run with different train-val-test splits for 5 times to get the average scores. |
| Hardware Specification | Yes | Our experiments are conducted on an A100 platform. All experiments were performed using an NVIDIA A100 80GB PCIe GPU, ensuring a consistent computational environment. |
| Software Dependencies | No | For our SBSN, we use Py Torch to implement it. |
| Experiment Setup | Yes | For a fair comparison, we set all the node embedding dimensions to 32, which is the same as that in SBGNN [9] and SBGCL [32], for all embedding-based methods. For other parameters in baselines, we follow the recommended settings in their original papers. We run up to 12,800,000 subgraph samples with K = 4 and T = 800 for SBSN for training and choose the model that performs the best in AUC metrics on the validation set. For our SBSN, we use Py Torch to implement it. We use the Adam optimizer with an initial learning rate of 0.002 and a weight decay of 1e-5. We set the maximum number of nodes in the topological subgraph to 800. |