Clustering then Propagation: Select Better Anchors for Knowledge Graph Embedding

Authors: KE LIANG, Yue Liu, Hao Li, Lingyuan Meng, Suyuan Liu, Siwei Wang, sihang zhou, Xinwang Liu

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on six datasets show that Rec Piece achieves higher performances but comparable or even fewer parameters compared to previous anchor-based KGE models, indicating that our model can select better anchors in a more scalable way.
Researcher Affiliation Academia 1National University of Defense Technology, Changsha, China 2Academy of Military Sciences, Beijing, China
Pseudocode No The paper describes the methodology in detail, but it does not include a structured pseudocode block or an algorithm labeled as such.
Open Source Code No The authors mention that the source code will be released after the double-blind review.
Open Datasets Yes Six benchmark datasets are leveraged to evaluate our Rec Piece as same as previous works do [13, 20, 30]. Specifically, FB15k-237 [64], WN18RR [18], Co DEx-L [54], and YAGO3-10 [45] are used for link prediction. The entity classification is carried out on two subsets (5% and 10% labeled) from WD50K [21], and the OGB WIKIKG 2 [23, 20] is the larger KGs for scalability analysis.
Dataset Splits Yes Datasets Tasks #Entities #Relationship #Edges #Train #Validation #Test FB15K-237 LP 14,505 237 310,079 272,115 17,526 20,438
Hardware Specification Yes All experiments are conducted on the server with 4-core Intel(R) Xeon(R) Platinum 8358 CPUs @ 2.60GHZ, a single 80 GB A100 GPU and 64GB RAM with Py Torch [49] libraries.
Software Dependencies No The paper mentions 'Py Torch [49] libraries' but does not specify a version number for PyTorch or any other relevant software libraries or dependencies.
Experiment Setup Yes The p( ) for feature preparation is selected as pretrained Node Piece [63] in the first few epochs. Besides, k-means [41, 44] is selected as g( ) for clustering, and the cluster number is set as '#Rel.' in Tab. 9 for different datasets. For a fair comparison, we set anchor numbers k for each dataset as the same as [20], and 2-layer MLP is adopted as f( ) feature propagation.