KAA: Kolmogorov-Arnold Attention for Enhancing Attentive Graph Neural Networks
Authors: Taoran Fang, Tianhong Gao, Chunping Wang, YihaoShang, Wei Chow, Lei CHEN, Yang Yang
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on both node-level and graph-level tasks using various backbone models show that KAA-enhanced scoring functions consistently outperform their original counterparts, achieving performance improvements of over 20% in some cases. |
| Researcher Affiliation | Collaboration | Taoran Fang Zhejiang University EMAIL Tianhong Gao Shanghai Jiaotong University EMAIL Chunping Wang Fin Volution Group EMAIL Yihao Shang Zhejiang University shang EMAIL Wei Chow Zhejiang University EMAIL Lei Chen Fin Volution Group EMAIL Yang Yang Zhejiang University EMAIL |
| Pseudocode | No | The paper describes algorithms and methods using mathematical formulas and descriptions, but it does not contain any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Our code is available at https://github.com/zjunet/KAA. |
| Open Datasets | Yes | For the node classification task, we select four citation network datasets (Sen et al., 2008; Hu et al., 2020) of varying scales: Cora, Cite Seer, Pub Med, and ogbn-arxiv, along with two product network datasets (Shchur et al., 2018), Amazon-Computers and Amazon-Photo. [...] For the graph classification task, we select four datasets (Ivanov et al., 2019; Zitnik & Leskovec, 2017) from bioinformatics and cheminformatics: PPI, MUTAG, ENZYMES, and PROTEINS. For the graph regression tasks, we select two datasets: ZINC (G omez-Bombarelli et al., 2018) and QM9 (Wu et al., 2018). |
| Dataset Splits | Yes | Cora 2708 5429 1433 7 140 / 500 / 1000 Cite Seer 3327 4732 3703 6 120 / 500 / 1000 Pub Med 19717 44338 500 3 60 / 500 / 1000 ogbn-arxiv 169343 1166243 128 40 90941 / 29799 / 48603 Amazon-Computers 13752 491722 767 10 10% / 10% / 80% Amazon-Photo 7650 238162 745 8 10% / 10% / 80% (from Table 4). PPI 24 2414 33838 50 121 80% / 10% / 10% MUTAG 188 17.93 19.76 7 2 80% / 10% / 10% ENZYMES 600 32.82 62.60 3 6 80% / 10% / 10% PROTEINS 1113 39.72 74.04 3 2 80% / 10% / 10% ZINC 12000 23.15 24.91 1 1 10000 / 1000 / 1000 QM9 130831 18.03 18.66 11 12 80% / 10% / 10% (from Table 5). |
| Hardware Specification | No | The paper does not provide specific details about the hardware used for running its experiments, such as GPU or CPU models. |
| Software Dependencies | No | The paper mentions software components like 'Adam optimizer' and 'B-spline functions' but does not provide specific version numbers for any libraries, frameworks, or programming languages. |
| Experiment Setup | Yes | For all models, we apply the dropout technique with dropout rates selected from [0, 0.1, 0.3, 0.5, 0.8]. Additionally, we utilize the Adam optimizer, choosing learning rates from [10-3, 5*10-3, 10-2] and weight decay values from [0, 5*10-4]. Regarding model architecture, the number of GNN layers is selected from [2, 3, 4, 5], the hidden dimension from [8, 16, 32, 64, 128, 256], and the number of heads from [1, 2, 4, 8]. For the KAN and MLP modules, we adopt an equal-width structure across all models, with the number of layers chosen from [2, 3, 4]. In the KAN modules, B-spline functions serve as the base functions, with the grid size selected from [1, 2, 4, 8] and the spline order chosen from [1, 2, 3]. We conduct five rounds of experiments with different random seeds for each setting and report the average results. |