TransGate: Knowledge Graph Embedding with Shared Gate Structure
Authors: Jun Yuan, Neng Gao, Ji Xiang3100-3107
AAAI 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct extensive experiments on link prediction and triplets classification. Experiments show that Trans Gate not only outperforms state-of-art baselines, but also reduces parameters greatly. |
| Researcher Affiliation | Academia | Jun Yuan,1,2,3 Neng Gao,1,2 Ji Xiang1,2 1Institute of Information Engineering, Chinese Academy of Sciences, Beijing, China 2State Key Laboratory of Information Security, Chinese Academy of Sciences, Beijing, China 3School of Cyber Security, University of Chinese Academy of Sciences, Beijing, China {yuanjun,gaoneng,xiangji}@iie.ac.cn |
| Pseudocode | No | The paper describes the Trans Gate architecture and training method using mathematical equations and textual descriptions, but it does not include formal pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not provide an explicit statement or link for the release of its own source code. |
| Open Datasets | Yes | Link prediction and triplets classification are implemented on two large-scale knowledge graphs: Word Net (Miller 1995) and Freebase (Bollacker et al. 2008). Word Net provides semantic knowledge of words. ... In this paper, we employ two data sets from Word Net: WN11 (Socher et al. 2013) and WN18RR (Dettmers et al. 2018). ... In this paper, we employ three data sets from Freebase: FB15K (Bordes et al. 2013), FB13 (Socher et al. 2013) and FB15K-237 (Toutanova and Chen 2015). |
| Dataset Splits | Yes | We select the hyper parameters of Trans Gate via grid search according to the Hits@10 on the validation set. ... The training process is stopped based on model s performance on the validation set. ... The optimal configurations are determined by the validation set. Table 3 lists statistics of the five data sets. |
| Hardware Specification | Yes | Additionally, we implemented our methods and Trans E in Tensor Flow (Abadi et al. 2016). Note that all models were run on a standard hardware of Inter(R) Core(TM) i7 2.6GHz + Ge Force GTX 960M, with the same mini-batch size. |
| Software Dependencies | No | The paper mentions implementing methods in TensorFlow but does not specify a version number for TensorFlow or any other software dependencies. |
| Experiment Setup | Yes | We search the learning rate α for Adam among {0.01, 0.1, 0.5}, the margin γ among {1, 2, 3, 4, 5, 6}, the embedding dimension m among {32, 50, 100, 200}, and the batch size B among {120, 480, 1440}. ... The best configurations are as follow: on WN18RR, γ = 5, α = 0.1, m = 200, B = 120 and taking L1 dissimilarity; on FB15K, γ = 1, α = 0.1, m = 200, B = 480 and taking L1 dissimilarity; on FB15K-237, γ = 4, α = 0.5, m = 200, B = 480 and taking L1 dissimilarity. |