Cycle Representation Learning for Inductive Relation Prediction
Authors: Zuoyu Yan, Tengfei Ma, Liangcai Gao, Zhi Tang, Chao Chen
ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our method achieves state-of-the-art performance on benchmarks. ... We compare our methods with state-of-the-art (SOTA) inductive relation prediction models on popular benchmark datasets. We also use ablation studies to demonstrate the efficacy of different proposed modules in our method. |
| Researcher Affiliation | Collaboration | 1Wangxuan Institute of Computer Technology, Peking University, Beijing, China 2T. J. Watson Research Center, IBM, New York, USA 3Department of Biomedical Informatics, Stony Brook University, New York, USA. |
| Pseudocode | No | No pseudocode or algorithm block found. |
| Open Source Code | Yes | The code is provided in https://github.com/pkuyzy/CBGNN. |
| Open Datasets | Yes | The datasets used in our settings are subsets of KG WN18RR (Toutanova & Chen, 2015), FB15k-237 (Dettmers et al., 2018), and NELL-995 (Xiong et al., 2017). |
| Dataset Splits | Yes | To evaluate the robustness of models, Teru et al. (2020) sample four different pairs of training sets and test sets with the increasing number of nodes and edges. The details of the benchmark datasets are shown in Table 5. ... We follow the settings in (Teru et al., 2020; Mai et al., 2021), that is, to view all the existing triplets in KG as positive triplets and sample negative triplets by replacing the head (or tail) of the triplet with a uniformly sampled random entity. ... we use the inductive training set as the validation set (while the training set and the test set are the same with (Teru et al., 2020; Mai et al., 2021)). |
| Hardware Specification | Yes | We run all the baseline methods with a cluster of two Intel Xeon Gold 5128 processors, 192GB RAM, and one GeForce RTX 2080 Ti graphics card. |
| Software Dependencies | No | No specific software versions (e.g., PyTorch 1.9, Python 3.8) are mentioned. |
| Experiment Setup | Yes | We adopt a 2-layer BR-LSTM to generate feature vectors for all the cycles in a cycle basis. Its output feature vector dimension is set to 20. A 2-layer GCN (Kipf & Welling, 2017) is adopted for the message passing of cycle basis, where ReLU serves as the activation function between GCN layers. We combine 20 different shortest path trees to learn the good rules in the given dataset. In the cycle graph, we select the top 2 most related cycles for each cycle. For all the modules, Adam is used as the optimizer, the dropout is set to 0.2, the epoch is set to 100 with an early-stopping of 20, the learning rate is 0.005 and the weight decay is 5e-5. |