Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Graph Few-Shot Learning via Adaptive Spectrum Experts and Cross-Set Distribution Calibration
Authors: Yonghao Liu, Yajun Wang, Chunli Guo, Wei Pang, Ximing Li, Fausto Giunchiglia, Xiaoyue Feng, Renchu Guan
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirically, GRACE consistently outperforms state-of-the-art baselines across a wide range of experimental settings. Our code can be found here. 6 Experiments To empirically validate the effectiveness of our proposed model, we utilize several widely adopted datasets for FSNC tasks, including Cora [23], Cite Seer [23], Amazon-Computer [47], Coauthor CS [47], DBLP [48], Cora Full [49], and, a large-scale dataset, ogbn-arxiv [50]. |
| Researcher Affiliation | Academia | 1Key Laboratory of Symbolic Computation and Knowledge Engineering of the Ministry of Education, College of Computer Science and Technology, Jilin University 2College of Software, Jilin University 3School of Mathematical and Computer Sciences, Heriot-Watt University 4RIKEN Center for Advanced Intelligence Project 5Department of Information Engineering and Computer Science, University of Trento |
| Pseudocode | Yes | A.1 Training Procedure Algorithm 1 Training Procedure of GRACE Require: A graph H = {V, E, X, A}. Ensure: The trained GRACE. 1: Perform the low-pass expert using Eq.1. 2: Perform the high-pass expert using Eqs.2 and 3. 3: Intergrate the outputs of low-pass and high-pass experts using Eq.4. 4: Perform cross-set distribution calibration using Eqs.5 and 6. 5: Optimize the proposed model by minimizing the loss in Eq.7. 6: Evaluate the model performance using query set with Eq.8. We present the detailed training procedure of GRACE in Algorithm 1. |
| Open Source Code | Yes | Empirically, GRACE consistently outperforms state-of-the-art baselines across a wide range of experimental settings. Our code can be found here. Our data and code are available at (https://anonymous.4open.science/ r/GRACE-7E41) |
| Open Datasets | Yes | To empirically validate the effectiveness of our proposed model, we utilize several widely adopted datasets for FSNC tasks, including Cora [23], Cite Seer [23], Amazon-Computer [47], Coauthor CS [47], DBLP [48], Cora Full [49], and, a large-scale dataset, ogbn-arxiv [50]. The statistics of these datasets are presented in Table 1. We present detailed descriptions of these adopted datasets in Appendix A.4. |
| Dataset Splits | Yes | Each dataset is divided into disjoint class sets for meta-training, meta-validation, and meta-testing. The details are as follows: Cora [4]: A citation graph where nodes represent academic papers and edges indicate citation relationships. Each node is assigned a label based on the paper s research topic. We divide the dataset into 3, 2, and 2 classes for meta-training, meta-validation, and meta-testing, respectively. ... For evaluation, we randomly generate multiple meta-testing tasks from the test set. Specifically, 100 tasks are sampled per evaluation, with each task comprising 10 query samples. |
| Hardware Specification | Yes | All experiments are carried out on an NVIDIA 3090Ti GPU to maintain consistent computational conditions and reproducibility. |
| Software Dependencies | No | During training, we use the Adam optimizer [56] with an initial learning rate of 0.001. For evaluation, we randomly generate multiple meta-testing tasks from the test set. Specifically, 100 tasks are sampled per evaluation, with each task comprising 10 query samples. To ensure the fairness and stability of our results, we conduct 5 independent experiments and report the average accuracy, standard deviation, and 95% confidence interval across these runs. |
| Experiment Setup | Yes | In the stage of adaptive spectrum experts, the low-pass expert is implemented using a two-layer GCN, with each layer followed by batch normalization and a Re LU activation. For the high-pass expert, the dimensions of all projection matrices are uniformly set to 32, i.e., d = 32. The gating network is implemented as a two-layer fully connected network, with the hidden dimension of 96. The temperature τ in Eq.4 is set to 2. Additionally, in the cross-set distribution calibration stage, the Gaussian kernel bandwidth σ is set to 1. During training, we use the Adam optimizer [56] with an initial learning rate of 0.001. |