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..

DyG-Mamba: Continuous State Space Modeling on Dynamic Graphs

Authors: Dongyuan Li, Shiyin Tan, Ying Zhang, Ming Jin, Shirui Pan, Manabu Okumura, Renhe Jiang

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Through exhaustive experiments on 12 datasets covering dynamic link prediction and node classification tasks, we show that Dy G-Mamba achieves state-of-the-art performance on most datasets, while demonstrating significantly improved computational and memory efficiency. Code is available at [https://github.com/Clearloveyuan/Dy G-Mamba].
Researcher Affiliation Academia 1The University of Tokyo, 2Institute of Science Tokyo, 3RIKEN Center for Advanced Intelligence Project, 4Griffith University EMAIL, EMAIL, EMAIL, EMAIL, EMAIL, EMAIL, EMAIL
Pseudocode Yes Pseudocodes are in Appendix C. Algorithm 1 Continuous SSM, Algorithm 2 Dynamic Link Prediction, Algorithm 3 Dynamic Node Classification
Open Source Code Yes Code is available at [https://github.com/Clearloveyuan/Dy G-Mamba]. We provide source code for reproducibility: https://anonymous/Dy GMamba. All implementation details could be accessed at the link: https://anonymous.4open/Dy GMamba.
Open Datasets Yes We evaluate our methods on a diverse set of dynamic graph datasets, including twelve publicly available datasets collected by Edgebank [30], which are publicly available1. We present the statistics of the datasets in Table 8... 1https://zenodo.org/records/dynamic-graphs
Dataset Splits Yes We evaluate performance on 12 datasets, each split into 70%/15%/15% for training, validation and testing.
Hardware Specification Yes Experiment Environment. We conduct experiments on an Ubuntu 22.04 LTS server equipped with one Intel(R) Core(TM) i9-10900X CPU @ 3.70GHz with 10 physical cores and NVIDIA RTX A6000 GPUs (48GB).
Software Dependencies Yes The code is written in Python 3.10 and we use Py Torch 2.1.0 on CUDA 11.8 to train the model.
Experiment Setup Yes We train each model for 100 epochs and select the best-performing checkpoint for testing. We repeat each experiment 10 times with different random seeds and report the mean and standard derivation. Details in Appendix D.2. For Dy G-Mamba, we list all configurations in Table 9. Learning rate 0.0001 Train Epochs 100 Optimizer Adam Dimension of time encoding d T 100 Dimension of co-occurrence d C 50 Dimension of aligned encoding d 50 Dimension of ti s encoder 4d 200 Dimension of output dout 172 Number of Mamba blocks 2 Dimension of SSM dssm 16 Expanded factor of Mamba 2 Number of Corss-Attention layer 1