Avoiding Structural Pitfalls: Self-Supervised Low-Rank Feature Tuning for Graph Test-Time Adaptation

Authors: Haoxiang Zhang, Zhuofeng Li, Qiannan Zhang, Ziyi Kou, Juncheng Li, Shichao Pei

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

Reproducibility Variable Result LLM Response
Research Type Experimental Through extensive experiments on six real-world datasets with diverse distribution shifts, we demonstrate that GOAT achieves consistent performance improvements across different pre-trained GNN backbones, outperforming state-of-the-art test-time adaptation methods.
Researcher Affiliation Academia Haoxiang Zhang EMAIL Halıcıoğlu Data Science Institute University of California, San Diego; Qiannan Zhang EMAIL Weill Cornell Medicine College Cornell University; Ziyi Kou EMAIL Department of Computer Science and Engineering University of Notre Dame; Juncheng Li EMAIL School of Computer Science and Technology East China Normal University; Shichao Pei EMAIL Department of Computer Science University of Massachusetts Boston
Pseudocode Yes Algorithm 1 GOAT for Full Test-Time Graph Adaptation
Open Source Code No In this section, we reveal the details of reproducing the results in the experiments. We will release the source code upon acceptance.
Open Datasets Yes We evaluate GOAT s performance on three types of distribution shifts across six benchmark datasets, following the experimental settings of EERM (Wu et al., 2023). The dataset statistics, along with a breakdown of these three distinct types of distribution shifts, are presented in Table 3: (1) Artificial Transformation for Cora (Yang et al., 2016) and Amazon-Photo (Shchur et al., 2018), where node features are replaced by synthetic features. (2) Cross-Domain transfers for Twitch-E (Rozemberczki et al., 2021) and FB-100 (Traud et al., 2012), involving graphs from different domains. (3) Temporal Evolution for Elliptic (Pareja et al., 2020) and OGB-Ar Xiv (Hu et al., 2020), utilizing dynamic datasets with natural evolving characteristics.
Dataset Splits Yes The datasets are split into training/validation/test sets with ratios of: 1/1/8 for Cora and Amazon-Photo; 1/1/5 for Twitch-E; 3/2/3 for FB-100; 5/5/33 for Elliptic; and 1/1/3 for OGBAr Xiv.
Hardware Specification Yes We perform experiments on NVIDIA Ge Force RTX 3090 GPUs. The GPU memory and running time reported in Table 3 are measured on one single RTX 3090 GPU. Additionally, we use eight CPUs, with the model name Intel(R) Xeon(R) Silver 4210R CPU @ 2.40GHz.
Software Dependencies No The operating system utilized in our experiments was Ubuntu 22.04.3 LTS (codename jammy).
Experiment Setup Yes For the setup of backbone GNNs, we majorly followed EERM (Wu et al., 2023): (a) GCN: the architecture setup is 5 layers with 32 hidden units for Elliptic and OGB-Ar Xiv, and 2 layers with 32 hidden units for other datasets, with batch normalization for all datasets. The pre-train learning rate is set to 0.001 for Cora and Amz-Photo, 0.01 for other datasets; the weight decay is set to 0 for Elliptic and OGB-Ar Xiv, and 0.001 for other datasets. ... For GOAT , we adopt Drop Edge as the augmentation function DE( ) and set the drop ratio to 0.05, K-layer aggregation in LRA set to 1 due to some GNN only has two layers in some datasets while the last GNN layer performs as a classifier head. We use Adam Optimizer for LRA module tuning. We further search the learning rate η in [1e-2, 5e-3, 1e-3, 5e-4, 1e-4, 5e-5, 1e-5, 1e-6] for different backbones, the virtual nodes number |n| in [1 , 2 , 5 , 10 , 20 ] of the class number C, the attention dim dattn in LRA in [2, 4, 8, 16, 32], total epochs T in [50, 100], and the patience in [1, 0.5, 0.1, 5e-2, 1e-2, 1e-3]. In the optimization target, we search the λ in [1, 3, 5, 10] and the α in [0.999, 0.9, 0.75, 0.5, 0.25, 0.1, 5e-2, 1e-2, 5e-3].