Relative and Absolute Location Embedding for Few-Shot Node Classification on Graph
Authors: Zemin Liu, Yuan Fang, Chenghao Liu, Steven C.H. Hoi4267-4275
AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on three public datasets demonstrate the state-of-the-art performance of RALE. |
| Researcher Affiliation | Collaboration | Zemin Liu,1 Yuan Fang,1 Chenghao Liu,1,2 Steven C.H. Hoi1,2 1 Singapore Management University, Singapore 2 Salesforce Research Asia, Singapore {zmliu, yfang}@smu.edu.sg, {chenghao.liu, shoi}@salesforce.com |
| Pseudocode | Yes | The pseudocode and complexity analysis are included in the supplementary. |
| Open Source Code | No | The paper does not provide a direct link or explicit statement about the release of its source code. |
| Open Datasets | Yes | We employ three public real-world datasets in our experiments. (1) Amazon (Hou et al. 2020) is an ecommerce network... (2) Email (Yin et al. 2017) is a communication network... (3) Reddit (Hamilton, Ying, and Leskovec 2017) is a social network... |
| Dataset Splits | Yes | We randomly split the node classes into training (i.e., the base classes), validation and testing (i.e., the novel classes), as presented in Table 1. ... Table 1: Summary of datasets. Nodes Edges Features Classes (Train/Val/Test) Amazon 13,381 245,778 767 10 (5/2/3) Email 909 13,733 128 28 (15/6/7) Reddit 231,371 11,606,876 602 41 (25/6/10) |
| Hardware Specification | No | The paper does not specify the hardware used for experiments (e.g., GPU/CPU models, memory). |
| Software Dependencies | No | The paper mentions software components like GNNs, Graph SAGE, GAT, and MAML, but does not specify their version numbers or other software dependencies. |
| Experiment Setup | Yes | Parameters and settings. For each method, we tune and select their hyperparameters and settings based on validation and guidance from the literature. For all GNNs and GNN+ s, we employ two layers, and further tune the dropout probability to 0.6, the dimension of the hidden layers to 32, as well as the learning rates for base class training and novel class updating or fine-tuning to 0.001 and 0.01, respectively. For specific architectures, we use 8 heads for the multi-head attention mechanism in GAT, and apply the mean aggregator for Graph SAGE. For all meta-learning methods, we use the same parameters in the above for the base GNN architecture, with a metalearning rate of 0.001. We also tune the dropout rate, and find that Meta-GNN and Proto-GNN is optimal when no dropout is used, whereas for RALE we use the dropout rate of 0.4 on the Amazon dataset and 0.5 on others. For MAML-based approaches, we further tune the number of gradient updates when adapting to the support set, which is set to 2 for Meta GNN and 1 for RALE, as well as the learning rate of adaptation α, which is set to 5.0 for both on the Amazon and Reddit datasets, and 0.2 on the Email dataset (α = 0.5 is applied in Email 1-shot setting for RALE). Finally, we report the default settings for hubs and random walks in our model RALE. We rank all nodes by their Page Rank scores in a descending order, and choose the top 5% nodes as hubs. The 5% here is called the hub ratio, i.e., |H|/|V|. To sample paths for a given task, we perform w = 200 random walks of length l = 50 starting from each node in the task (Perozzi, Al-Rfou, and Skiena 2014). A sliding window of lp/2 is applied on each sampled walk to extract path segments. Segments ending with a hub are further joined to form paths up to length lp = 6. |