A Scalable and Effective Alternative to Graph Transformers

Authors: Kaan Sancak, Zhigang Hua, Jin Fang, Yan Xie, Andrey Malevich, Bo Long, Muhammed Fatih Balin, Ümit V. Çatalyürek

AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our study on synthetic datasets reveals that GECO reaches 169 speedup on a graph with 2M nodes w.r.t. optimized attention. Further evaluations on diverse range of benchmarks showcase that GECO scales to large graphs where traditional GTs often face memory and time limitations. Notably, GECO consistently achieves comparable or superior quality compared to baselines, improving the SOTA up to 4.5%, and offering a scalable and effective solution for large-scale graph learning. 4 Experiments
Researcher Affiliation Collaboration 1School of Computational Science and Engineering, Georgia Institute of Technology, Atlanta, GA, USA 2Meta AI
Pseudocode Yes Algorithm 1: Forward pass of GCB Operator; Algorithm 2: End-to-end GECO Model Training
Open Source Code Yes Code https://github.com/kaansancak/GECO
Open Datasets Yes Long Range Graph Benchmark (LRGB). Table 1 presents our evaluation on the LRGB, a collection of graph tasks designed to test a model s ability to capture long-range dependencies. PCQM4Mv2. Table 2 shows that GECO outperforms both GNN and GT baselines on PCQM4Mv2 in prediction quality. Table 3: Accuracy on large node prediction datasets: the first, second, and third are highlighted. We reuse the results from (Han et al. 2023; Shirzad et al. 2023; Zeng et al. 2021), and run Exphormer locally except Arxiv.
Dataset Splits No The paper mentions using a 'validation set' for evaluation on PCQM4Mv2 and states 'For dataset and hyperparameter details please refer to the extended version.' in Section 4.1. It does not explicitly provide specific dataset split information (percentages or counts) in the main text.
Hardware Specification Yes Even the most computation-intensive GTs, such as Graphormer, can be trained on these datasets using Nvidia-V100 (32GB) or Nvidia-A100 (40GB) GPUs (Ying et al. 2021; Rampasek et al. 2022).
Software Dependencies No The paper does not provide specific version numbers for ancillary software dependencies such as programming languages, libraries, or frameworks used in their implementation.
Experiment Setup No For dataset and hyperparameter details please refer to the extended version.