Graph Few-shot Learning with Task-specific Structures
Authors: Song Wang, Chen Chen, Jundong Li
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We further conduct extensive experiments on five node classification datasets under both singleand multiple-graph settings to validate the superiority of our framework over the state-of-the-art baselines. |
| Researcher Affiliation | Academia | Song Wang University of Virginia sw3wv@virginia.edu Chen Chen University of Virginia zrh6du@virginia.edu Jundong Li University of Virginia jundong@virginia.edu |
| Pseudocode | No | The paper describes its methods using prose and mathematical equations but does not include any explicit pseudocode or algorithm blocks. |
| Open Source Code | Yes | Our code is provided at https://github.com/Song W-SW/GLITTER. |
| Open Datasets | Yes | To evaluate the performance of GLITTER on the few-shot node classification task, we conduct experiments on five prevalent real-world graph datasets, two for the multiple-graph setting and three for the single-graph setting. The detailed statistics are provided in Table 1. For the multiple-graph setting, we use the following two datasets: (1) Tissue-PPI [49] consists of 24 protein-protein interaction (PPI) networks... (2) Fold-PPI [18] contains of 144 tissue PPI networks... For the single-graph setting, we leverage three datasets: (1) DBLP [30] is a citation network... (2) Cora-full [2] is also a citation network... (3) ogbn-arxiv [17] is a citation network... |
| Dataset Splits | Yes | Specifically, Ti = {Si, Qi}, where Si is the support set of Ti and consists of K labeled nodes for each of N classes (i.e., |Si| = NK). The corresponding label set of Ti is Yi, where |Yi| = N. Yi is sampled from the whole training label set Ytrain. With Si as references, the model is required to classify nodes in the query set Qi, which contains Q unlabeled samples. After training, the model will be evaluated on a series of meta-test tasks, which follow a similar setting as meta-training tasks. The training process of our framework is conducted on a series of meta-training tasks. After training, the model will be evaluated on a specific number of meta-test tasks. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., GPU/CPU models, memory) used for running its experiments. |
| Software Dependencies | No | The paper mentions 'PyTorch' in its references but does not specify version numbers for any software dependencies used in its implementation. |
| Experiment Setup | Yes | Specifically, for each meta-task, we first perform several update steps to learn task-specific structures based on the two structure losses: θ(i) S = θ(i 1) S α L(i) S , where θS denotes the parameters used to learn task-specific structures, and L(i) S = L(i) N +L(i) M . θ(i) G = θ(i 1) G α L(i) support, where L(i) support = P j yk,j log p(i) k,j is the cross-entropy loss, and α is the base learning rate. After repeating these steps for η times, the loss on the query set will be used for the meta-update: θG = θ(η) G β1 L(η) query and θS = θ(η) S β2 L(η) S , where β1 and β2 are the meta-learning rates, and L(η) query is the cross-entropy loss calculated on query nodes. |