Comprehensive Analysis of Negative Sampling in Knowledge Graph Representation Learning

Authors: Hidetaka Kamigaito, Katsuhiko Hayashi

ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our empirical analysis on the FB15k-237, WN18RR, and YAGO3-10 datasets showed that the results of actually trained models agree with our theoretical findings.
Researcher Affiliation Academia 1Nara Institute of Science and Technology (NAIST), Nara, Japan 2Hokkaido University, Hokkaido, Japan.
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code Yes Our code is available at https://github.com/kamigaito/icml2022.
Open Datasets Yes We used FB15k-237 (Toutanova & Chen, 2015), WN18RR, and YAGO3-10 (Dettmers et al., 2018) as the datasets.
Dataset Splits Yes Table 1: Statistics for each dataset. ... FB15k-237 ... Train 272,115 Valid 17,535 Test 20,466 ... WN18RR ... Train 86,835 Valid 3,034 Test 3,134 ... YAGO3-10 ... Train 1,079,040 Valid 4,978 Test 4,982
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory) used for running the experiments. It only mentions using Adam as an optimizer.
Software Dependencies No The paper mentions using Adam (Kingma & Ba, 2015) as an optimizer and refers to implementations from previous works, but it does not specify software dependencies with version numbers.
Experiment Setup Yes Settings: ... We chose margin term γ from {0.0, 3.0, 6.0, 9.0, 9.58} on FB15k-237 and {0.0, 2.0, 4.0, 6.0, 10.62} on WN18RR for each initial learning rate in {e-2, e-3, 5e-5}. ... We varied the number of negative samples ν from {32, 64, 128, 256, 512} on FB15k-237 and {64, 128, 256, 512, 1024} on WN18RR for each initial learning rate in {e-2, e-3, 5e-5}. ... Table 6: Hyperparameters for each model. Batch denotes the batch size, Dim denotes the hidden dimension size, α denotes the temperature parameter for SANS, and Step denotes the max steps in training. ... Table 7: Hyperparameters for each model in 4.5. LR denotes the learning rate.