How to Turn Your Knowledge Graph Embeddings into Generative Models
Authors: Lorenzo Loconte, Nicola Di Mauro, Robert Peharz, Antonio Vergari
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 7 Empirical Evaluation We aim to answer the following research questions: RQ1) are Ge KCs competitive with commonly used KGEs for link prediction? RQ2) Does integrating domain constraints in Ge KCs benefit training and prediction?; RQ3) how good are the triples sampled from Ge KCs? and Table 1: Ge KCs are competitive with their energy-based counterparts. Best average test MRRs of CP, COMPLEX and Ge KCs trained with the PLL and MLE objectives (Eqs. (1) and (2)). For standard deviations and training times see Table F.2. |
| Researcher Affiliation | Academia | Lorenzo Loconte University of Edinburgh, UK l.loconte@sms.ed.ac.uk Nicola Di Mauro University of Bari, Italy nicola.dimauro@uniba.it Robert Peharz TU Graz, Austria robert.peharz@tugraz.at Antonio Vergari University of Edinburgh, UK avergari@ed.ac.uk |
| Pseudocode | No | No section or figure explicitly labeled 'Pseudocode' or 'Algorithm'. |
| Open Source Code | Yes | Code is available at https://github.com/april-tools/gekcs. |
| Open Datasets | Yes | We evaluate Ge KCs on standard KG benchmarks for link prediction4: FB15k-237 [62], WN18RR [21] and ogbl-biokg [32], whose statistics can be found in Appendix F.1. |
| Dataset Splits | Yes | The models are trained until the MRR computed on the validation set does not improve after three consecutive epochs. and Table F.1: Dataset statistics. Statistics of multi-relational knowledge graphs: number of entities (|E|), number of predicates (|R|), number of training/validation/test triples. |
| Hardware Specification | Yes | Hardware. Experiments on the smaller knowledge graphs FB15K-237 and WN18RR were run on a single Nvidia GTX 1060 with 6 Gi B of memory, while those on the larger ogbl-biokg and ogbl-wikikg2 were run on a single Nvidia RTX A6000 with 48 Gi B of memory. |
| Software Dependencies | No | No specific software dependencies with version numbers (e.g., Python, PyTorch) are listed. Only general software like 'Adam' as an optimizer is mentioned without a version. |
| Experiment Setup | Yes | Hyperparameters. All models are trained by gradient descent with either the PLL or the MLE objective (Eqs. (1) and (2)). We set the weights ωs, ωr, ωo of the PLL objective all to one... The models are trained until the MRR computed on the validation set does not improve after three consecutive epochs. We fix the embedding size d = 1000 for both CP and COMPLEX and use Adam [38] as optimiser with 10 3 as learning rate. An exception is made for Ge KCs obtained via non-negative restriction (Section 4.1), for which a learning rate of 10 2 is needed... We search for the batch size in {5 102, 103, 2 103, 5 103} based on the validation MRR. |