A Topology-aware Graph Coarsening Framework for Continual Graph Learning

Authors: Xiaoxue Han, Zhuo Feng, Yue Ning

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

Reproducibility Variable Result LLM Response
Research Type Experimental We empirically demonstrate that the learning process on the reduced graph can closely approximate that on the original graph. We compare TACO with a wide range of state-of-the-art baselines, proving its superiority and the necessity of preserving high-quality topological information for effective replaying. We conduct extensive experiments and perform comprehensive ablation studies to evaluate the effectiveness of TACO and Re Pro.
Researcher Affiliation Academia Xiaoxue Han Stevens Institute of Technology xhan26@stevens.edu Zhuo Feng Stevens Institute of Technology zfeng12@stevens.edu Yue Ning Stevens Institute of Technology yue.ning@stevens.edu
Pseudocode Yes The pseudocode of TACO is described in Algorithm 1 and Re Pro is described in Algorithm 2 in Appendix D.2.
Open Source Code Yes Our code is available at: https://github.com/hanxiaoxue114/TACO.
Open Datasets Yes We conduct experiments on time-stamped graph datasets: Kindle [14, 31], DBLP [44] and ACM [44] to evaluate the performance of TACO.
Dataset Splits Yes For each task, we randomly split all nodes into training, validation, and test sets with the ratio of 30/20/50.
Hardware Specification Yes All experiments were conducted on a NVIDIA GeForce RTX 3090 GPU.
Software Dependencies No The paper mentions using GNN models like GCN, GAT, and GIN as backbones, but it does not specify version numbers for any software libraries or dependencies (e.g., PyTorch, TensorFlow versions).
Experiment Setup Yes For each task, we randomly split all nodes into training, validation, and test sets with the ratio of 30/20/50. For the baseline CGL models the memory strengths are searched from {10i|i [ 5...5]} or {0.1, 0.2...0.9}. For TACO, by default the reduction ratio is 0.5; memory buffer size for Node Fidelity Preservation is 200; node degree is used to determine a node s importance score, and Reservoir Sampling is chosen as the node sampling strategy. We chose GCN and GAT as the GNN backbones. For both of them, we set the number of layers as 2 and the hidden layer size as 48. For GAT, we set the number of heads as 8.