Less is More: One-shot Subgraph Reasoning on Large-scale Knowledge Graphs
Authors: Zhanke Zhou, Yongqi Zhang, Jiangchao Yao, quanming yao, Bo Han
ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirically, we achieve promoted efficiency and leading performances on five large-scale benchmarks. The major experiments are conducted with Py Torch (Paszke et al., 2017) and one NVIDIA RTX 3090 GPU. We conduct extensive experiments on five large-scale datasets and achieve an average of 94.4% improvement in efficiency of prediction and 6.9% promotion in effectiveness of prediction (Sec. 5). |
| Researcher Affiliation | Academia | Zhanke Zhou1 Yongqi Zhang2 Jiangchao Yao3 Quanming Yao4 Bo Han1 1TMLR Group, Hong Kong Baptist University 2The Hong Kong University of Science and Technology (Guangzhou) 3CMIC, Shanghai Jiao Tong University 4Tsinghua University |
| Pseudocode | Yes | Algorithm 1 One-shot-subgraph Link Prediction on Knowledge Graphs |
| Open Source Code | Yes | The code is publicly available at: https://github.com/tmlr-group/one-shot-subgraph. |
| Open Datasets | Yes | We use five benchmarks with more than ten thousand entities (see Tab. 11), including WN18RR (Dettmers et al., 2017), NELL-995 (Xiong et al., 2017), YAGO3-10 (Suchanek et al., 2007), OGBL-BIOKG, and OGBL-WIKIKG2 (Hu et al., 2020). |
| Dataset Splits | Yes | Table 11: Statistics of the five KG datasets with more than ten-thousand entities. Fact triplets in E are used to build the graph, and Etrain, Eval, Etest are edge sets of training, validation, and test set. |
| Hardware Specification | Yes | The major experiments are conducted with Py Torch (Paszke et al., 2017) and one NVIDIA RTX 3090 GPU. The OGB datasets are run with one NVIDIA A100 GPU. |
| Software Dependencies | No | The paper mentions 'Py Torch (Paszke et al., 2017)' but does not provide a specific version number for PyTorch or any other software dependency. |
| Experiment Setup | Yes | The ranges for design dimensions of the configuration space are shown below, where the upper is intra-layer design while the lower is inter-layer design. DROPOUT( ) ACT( ) AGG( ) MESS( ) Dimension (0, 0.5) Identity, Relu, Tanh Max, Mean, Sum MDRUM, MNBFNet, MREDGNN 16, 32, 64, 128 No. layers (L) Repre. initialization Layer-wise shortcut Repre. concatenation READOUT( ) {4, 6, 8, 10} Binary, Relational True, False True, False Linear, Dot product. |