Optimal Embedding Guided Negative Sample Generation for Knowledge Graph Link Prediction
Authors: Makoto Takamoto, Daniel Onoro Rubio, Wiem Ben Rim, Takashi Maruyama, Bhushan Kotnis
TMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | To evaluate its efficacy, we conducted comprehensive experiments across multiple datasets. The results consistently demonstrate significant improvements in link prediction performance across various KGE models and negative sampling methods. |
| Researcher Affiliation | Collaboration | Makoto Takamoto EMAIL NEC Laboratories Europe, Heidelberg, Germany Daniel Oõro-Rubio EMAIL NEC Laboratories Europe, Heidelberg, Germany Wiem Ben Rim University College London, London, UK Takashi Maruyama NEC Laboratories Europe, Heidelberg, Germany Bhushan Kotnis Coresystems AG, Zurich, Switzerland |
| Pseudocode | No | The paper includes theoretical derivations and equations but does not feature any explicitly labeled pseudocode or algorithm blocks. The methods are described textually or through mathematical formulas. |
| Open Source Code | Yes | An implementation of the method and experiments are available at https://github.com/nec-research/EMU-KG. |
| Open Datasets | Yes | Furthermore, we evaluate them on three widely used knowledge graphs, namely FB15k-237 (Toutanova & Chen, 2015), WN18RR (Dettmers et al., 2018), and YAGO3-10 (Mahdisoltani et al., 2013) whose detailed statistics are provided in Appendix F. |
| Dataset Splits | No | The paper mentions that "training, validation and testing refer to the number of triples under each split" in Appendix F, but Table 4 only provides the total number of triples for each dataset without specifying the actual percentages or counts for the training, validation, and test splits. |
| Hardware Specification | Yes | All the experiments other than HAKE were performed on one Nvidia Ge Force GTX 1080 Ti GPU for each run. The experiments with HAKE were performed on one Nvidia Ge Force RTX 3090 GPU for each run. All experiments were performed on a single NVIDIA A100 GPU. |
| Software Dependencies | Yes | The models were implemented by Py Torch 2.1.0 with CUDA11.8. |
| Experiment Setup | Yes | The optimization was performed using Adam (Kingma & Ba, 2014) for 105 iterations with 256 negative samples. The hyper-parameter tuning was performed with Optuna (Akiba et al., 2019). ... A more detailed hyper-parameters are provided in Table 5 and Table 6. We tuned our hyperparameters, including the learning rate and the coefficient for weight-decay for baseline scores, through 10000 iterations on the FB15K-237 validation dataset using Optuna (Akiba et al., 2019). The hyperparameters for EMU are detailed in Table 7. |