INDIGO: GNN-Based Inductive Knowledge Graph Completion Using Pair-Wise Encoding
Authors: Shuwen Liu, Bernardo Grau, Ian Horrocks, Egor Kostylev
NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experiments show that our model outperforms state-of-the-art approaches on inductive KG completion benchmarks. and 4 Evaluation |
| Researcher Affiliation | Academia | 1Department of Computer Science, University of Oxford, UK {shuwen.liu, bernardo.cuenca.grau, ian.horrocks}@cs.ox.ac.uk 2Department of Informatics, University of Oslo egork@ifi.uio.no |
| Pseudocode | No | The paper describes the model and its components using text and mathematical formulas but does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | No | The paper states 'We have implemented our approach using Python and Py Torch v1.4.0 in a system called INDIGO.' but does not provide any information about open-sourcing the code or a link to a repository. |
| Open Datasets | Yes | We exploit a number of benchmarks proposed by Teru et al. [21] and Hamaguchi et al. [5] for inductive KG completion. The benchmarks by Teru et al., 12 in total, are based on transductive benchmarks FB15K-237 [22], NELL-995 [26], and WN18RR [4]... Similarly, the benchmarks by Hamaguchi et al., 9 in total, are all based on a transductive benchmark Word Net11 [19]... To address this limitation, we have designed a new benchmark, called INDIGOBM, which is based on FB15K-237 and has the same structure and assumptions as the existing benchmarks... |
| Dataset Splits | Yes | Each of these benchmarks provides the following: disjoint sets T and V of triples with Sig(V) Sig(T ) for training and validation; an incomplete KG Ktest and a set Λ+ test of test triples... and The training set T of a benchmark is first randomly split, with ratio 9:1, into an incomplete KG Ktrain and a set Λ+ train of triples assumed to hold in the completion of Ktrain. |
| Hardware Specification | Yes | All experiments were performed on an Intel(R) Xeon(R) machine with 8 cores and a 2.6 GHz CPU equipped with 540 GB of RAM running Fedora 33 (x86_64). |
| Software Dependencies | Yes | We have implemented our approach using Python and Py Torch v1.4.0 |
| Experiment Setup | Yes | We set as hyper-parameters the number of layers in the GCN (2, 3, or 4), the dimension of vectors in the hidden layers (32, 64, or 128) and the learning rate (0.01 or 0.001) and cross-validated them on each of the benchmarks using the validation sets to obtain a most favourable (for all benchmarks) setting of 2 layers, vector dimension of 64, and learning rate 0.001. Our INDIGO system is trained as a denoising autoencoder [23]. ... trained for 3,000 epochs using Adam optimisation with L2 penalty 5e-8. |