Lifelong Embedding Learning and Transfer for Growing Knowledge Graphs

Authors: Yuanning Cui, Yuxin Wang, Zequn Sun, Wenqiang Liu, Yiqiao Jiang, Kexin Han, Wei Hu

AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct experiments regarding link prediction accuracy, knowledge transfer capability, and learning efficiency to validate the proposed model, LKGE. The datasets and source code are available at https://github.com/nju-websoft/LKGE.
Researcher Affiliation Collaboration Yuanning Cui1, Yuxin Wang1, Zequn Sun1, Wenqiang Liu3, Yiqiao Jiang3, Kexin Han3, Wei Hu1,2* 1 State Key Laboratory for Novel Software Technology, Nanjing University, China 2 National Institute of Healthcare Data Science, Nanjing University, China 3 Interactive Entertainment Group, Tencent Inc, China
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code Yes The datasets and source code are available at https://github.com/nju-websoft/LKGE.
Open Datasets Yes To simulate a variety of aspects of KG growth, we create four datasets based on FB15K-237 (Toutanova and Chen 2015), which are entity-centric, relation-centric, fact-centric, and hybrid. We denote them by ENTITY, RELATION, FACT and HYBRID, respectively. ... The datasets and source code are available at https://github.com/nju-websoft/LKGE.
Dataset Splits Yes For each snapshot, we randomly divide the new fact set T i into a training set Di, a validation set Vi and a test set Qi by a split ratio of 3:1:1.
Hardware Specification No The paper does not provide any specific details about the hardware used to run the experiments (e.g., CPU, GPU models, memory).
Software Dependencies No The paper mentions 'Adam optimizer' and 'Trans E' but does not specify any software names with version numbers (e.g., Python, PyTorch, TensorFlow versions).
Experiment Setup Yes For a fair comparison, we first tune the hyperparameters of the base model using grid-search: learning rate in {0.0005, 0.0001, 0.001}, batch size in {1024, 2048}, embedding dimension in {100, 200}. Then, we use the same base model for LKGE and all competitors, and tune other hyperparameters. For the regularization models, the α of regularization loss is in {0.01, 0.1, 1.0}. For our model, the β of MAE loss is in {0.01, 0.1, 1.0}. For all competitors, we use Adam optimizer and set the patience of early stopping to 3.