Explainable GNN-Based Models over Knowledge Graphs
Authors: David Jaime Tena Cucala, Bernardo Cuenca Grau, Egor V. Kostylev, Boris Motik
ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate our approach by applying it to classification tasks in knowledge graph completion. We demonstrate the effectiveness of our approach by applying it to classification tasks in KG completion. Using well-known benchmarks, we compare our system s performance with the state-of-the-art KG completion systems DRUM (Sadeghian et al., 2019) and Any BURL (Meilicke et al., 2019). We show that, in addition to providing an exact correspondence between the model and the extracted rules, our technique also offers competitive performance. |
| Researcher Affiliation | Academia | David Tena Cucala Department of Computer Science University of Oxford, UK Bernardo Cuenca Grau Department of Computer Science University of Oxford, UK Egor V. Kostylev Department of Informatics University of Oslo, Norway Boris Motik Department of Computer Science University of Oxford, UK |
| Pseudocode | Yes | Algorithm 1 Rule extraction algorithm Input: m : natural number (max. num. of atoms in the bodies of extracted rules) M : an MGNN |
| Open Source Code | Yes | Our proofs are given in the appendix, and the source code is available as supplementary material. |
| Open Datasets | Yes | We used the 12 KG completion benchmarks by Teru et al. (2020), which are based on FB15K-237 (Bordes et al., 2013), NELL-995 (Xiong et al., 2017), and WN18RR (Dettmers et al., 2018). |
| Dataset Splits | Yes | Each benchmark provides disjoint datasets T , V, and S for training, validation, and testing, respectively; dataset statistics are shown in Table 1. The training dataset T was split with a 9:1 ratio into an incomplete dataset TI and a set TM of missing facts that should be added to TI. |
| Hardware Specification | Yes | We implemented the procedure in Pytorch Geometric v1.5.0, and we ran it on a laptop running mac OS 10.15.7 with 8 GB of RAM and an Intel Core i5 2.30 GHz CPU. |
| Software Dependencies | Yes | We implemented the procedure in Pytorch Geometric v1.5.0, and we ran it on a laptop running mac OS 10.15.7 with 8 GB of RAM and an Intel Core i5 2.30 GHz CPU. |
| Experiment Setup | Yes | We trained MGNNs with two layers and the Re LU activation function; we used cross-entropy loss with a logistic sigmoid on λL(v) as the output probability; finally, we used the Adam optimisation algorithm with the standard learning rate (0.01) and weight decay (5 10 4), and a maximum of 50, 000 epochs. |