What Matters in Graph Class Incremental Learning? An Information Preservation Perspective

Authors: Jialu Li, Yu Wang, Pengfei Zhu, Wanyu Lin, Qinghua Hu

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 5 Experiments Table 1: Performance comparison on Cora Full, Arxiv, and Reddit for GCIL setting. Table 3: Ablation comparisons of graph spatial information preservation.
Researcher Affiliation Academia 1College of Intelligence and Computing, Tianjin University, Tianjin, China 2Engineering Research Center of City Intelligence and Digital Governance, Ministry of Education of the People s Republic of China, Tianjin, China 3Haihe Lab of ITAI, Tianjin, China 4Department of Computing, The Hong Kong Polytechnic University, Hong Kong, China
Pseudocode Yes Algorithm 1 Framework of GSIP
Open Source Code Yes The code is available through https://github.com/Jillian555/GSIP.
Open Datasets Yes We utilize five public datasets to evaluate the effectiveness of the proposed method in GCIL, the statistics of datasets are reported in Appendix B.1. ... Cora Full [48], ... Arxiv [49] and Reddit [50], ... Cora [51] and Citeseer [51].
Dataset Splits Yes The train-validation-test splitting ratios are 60%, 20%, and 20% for all datasets.
Hardware Specification Yes Our model is deployed in Py Torch with an NVIDIA RTX 3090 GPU and trained with 200 epochs for every task.
Software Dependencies No Our model is deployed in Py Torch with an NVIDIA RTX 3090 GPU and trained with 200 epochs for every task. (Only mentions PyTorch without a version number.)
Experiment Setup Yes Our model is deployed in Py Torch with an NVIDIA RTX 3090 GPU and trained with 200 epochs for every task. We use Adam with weight decay for optimization, and the learning rate is set to 0.005. We use a two-layer GCN with a hidden dimension 256 as the backbone.