Exploring Relational Semantics for Inductive Knowledge Graph Completion
Authors: Changjian Wang, Xiaofei Zhou, Shirui Pan, Linhua Dong, Zeliang Song, Ying Sha4184-4192
AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results on benchmark datasets show that our model outperforms state-of-the-art models for inductive KGC. |
| Researcher Affiliation | Academia | 1 Institute of Information Engineering, Chinese Academy of Sciences, Beijing, China 2 School of Cyber Security, University of Chinese Academy of Sciences, Beijing, China 3 Faculty of Information Technology, Monash University, Melbourne, Australia 4 College of Informatics, Huazhong Agricultural University, Wuhan, China |
| Pseudocode | Yes | Algorithm 1 shows the whole training procedure of our model. We jointly optimize these two objectives according to Eq. (8) and Eq. (11). Finally, the optimized aggregator and generator will be used to represent the emerging entities and perform KGC task. Algorithm 1: Model Training Require: a knowledge graph K, the number of iterations T, the number of samples m, |R| prior distributions p = {pq | q R} Ensure: aggregator A, generator G 1: for iterator = 1 to T do 2: Sample m positive examples with corresponding relations {(zi, ri)}m i=1 from the prior distributions p; 3: Sample m entities S = {si}m i=1 from E; 4: Sample m relations {q i}m i=1 from R; 5: Update D by Eq. (8) to maxmize: 1 m Pm i=1 log D(zi, ri) + log(1 D(G(A(si), qi), qi) 6: Update G and A by Eq. (8) to minimize: 1 m Pm i=1 log(1 D(G(A(si), qi), qi) 7: Extract training examples { s, q, o | s S} from K; 8: Update G and A according to Eq. (11); 9: end for |
| Open Source Code | Yes | The code and datasets are available at https://github.com/changjianw/CFAG. |
| Open Datasets | Yes | To evaluate our proposed model, we adopt two benchmark datasets: FB15k-237 (Toutanova and Chen 2015) and NELL-995 (Xiong, Hoang, and Wang 2017). ... The original training set is split into the training set and the auxiliary set, where the training set does not contain the emerging entities and each triple in the auxiliary set contains only one emerging entity. |
| Dataset Splits | Yes | The validation set is constructed by removing triples which contain emerging entities from the original validation set. ... Best models are selected by using early stopping according to MRR on the validation sets. |
| Hardware Specification | No | The paper does not specify the hardware used for running the experiments (e.g., CPU, GPU models, memory, or cloud instances). |
| Software Dependencies | No | The paper mentions using 'Adam (Kingma and Ba 2015) optimizer' and 'ReLU as the activation function' but does not specify version numbers for any software libraries or dependencies. |
| Experiment Setup | Yes | We train our models using Adam (Kingma and Ba 2015) optimizer and use grid search to select the hyperparameters of our model. Hyperparameter ranges are as follows: learning rate lr in {0.1, 0.01, 0.005, 0.001}, α in Eq. (5) in {0, 0.3, 0.5, 0.7, 1.0}, embedding size d E in {100, 200, 300}, output size of CG-AGG d A in {200, 500, 1000}, the number of filters K in {10, 50, 100}. The optimal hyperparameter configurations are lr = 0.001, α = 0.5, d E = 100, d A = 200 for all datasets, K = 50 for FB15k-237-Sub and FB15k237-Obj, K = 10 for NELL-995-Sub and NELL-995-Obj. We use Re LU as the activation function and multiple multivariate Gaussian distributions as the prior distributions. |