Generalizing Knowledge Graph Embedding with Universal Orthogonal Parameterization
Authors: Rui Li, Chaozhuo Li, Yanming Shen, Zeyu Zhang, Xu Chen
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirically, Gold E achieves state-of-the-art performance on three standard benchmarks. Codes are available at https://github.com/xxrep/Gold E. To comprehensively validate the effectiveness of Gold E, we conduct extensive experiments on link prediction task, and attempt to answer the following questions: (1) How does Gold E perform on link prediction compared to existing KGE approaches? (Section 4.2) (2) What is the effect of configuring different geometries and dimensions for orthogonal relation transformations on the performance of Gold E? (Section 4.3) (3) From a fine-grained perspective, does Gold E effectively model different relation types? (Section 4.4) (4) Can Gold E still learn high-quality representations even when the embedding size k is restricted? (Section 4.5) |
| Researcher Affiliation | Academia | 1Gaoling School of Artificial Intelligence, Renmin University of China, Beijing, China 2Key Laboratory of Trustworthy Distributed Computing and Service (MOE), Beijing University of Posts and Telecommunications, China 3School of Computer Science and Technology, Dalian University of Technology, China. |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Codes are available at https://github.com/xxrep/Gold E. |
| Open Datasets | Yes | Datasets: We evaluate Gold E framework on three standard benchmarks: WN18RR (Dettmers et al., 2018), FB15k-237 (Toutanova & Chen, 2015) and YAGO3-10 (Mahdisoltani et al., 2015). Refer to Appendix H for statistical details. Table 8. Statistics of five standard benchmarks. Dataset #Entity #Relation #Training #Validation #Test WN18RR 40,943 11 86,835 3,034 3,134 FB15k-237 14,541 237 272,115 17,535 20,466 YAGO3-10 123,182 37 1,079,040 5,000 5,000 |
| Dataset Splits | Yes | Table 8. Statistics of five standard benchmarks. Dataset #Entity #Relation #Training #Validation #Test WN18RR 40,943 11 86,835 3,034 3,134 FB15k-237 14,541 237 272,115 17,535 20,466 YAGO3-10 123,182 37 1,079,040 5,000 5,000 |
| Hardware Specification | No | The paper mentions "Public Computing Cloud, Renmin University of China" in the acknowledgements, but it does not provide specific hardware details such as GPU models, CPU models, or memory specifications used for running the experiments. |
| Software Dependencies | No | The paper mentions using "Adam (Kingma & Ba, 2015) as the optimizer" but does not specify version numbers for Adam or any other software libraries or dependencies. Therefore, specific ancillary software details with version numbers are not provided. |
| Experiment Setup | Yes | Implementation details: For fair comparisons, we follow (Li et al., 2022) to fix the embedding dimension k, ensuring that the number of parameters is comparable to the baselines. Specifically, we follow (Li et al., 2022) to fix the embedding size k of each entity as 800, 600, 1000 on WN18RR, FB15k-237 and YAGO3-10 datasets, respectively. Hyperparameters: We use Adam (Kingma & Ba, 2015) as the optimizer and fine-tune the hyperparameters on the validation dataset. The hyperparameters are tuned by the random search (Bergstra & Bengio, 2012), including batch size b, self-adversarial sampling temperature α, fixed margin γ, learning rate lr, dimension k , number of elliptic component spaces m P, and number of hyperbolic component spaces m Q. The hyperparameter search space is shown in Table 10. Table 10: Hyperparameter Search Space (lists specific choices for b, γ, lr, k, m P, m Q and ranges for α). |