A Theory of Link Prediction via Relational Weisfeiler-Leman on Knowledge Graphs
Authors: Xingyue Huang, Miguel Romero, Ismail Ceylan, Pablo Barceló
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct an experimental analysis to verify the impact of various model choices, particularly pertaining to initialization, history, message computation, and global readout functions. We also conduct both inductive and transductive experiments on various real-world datasets, empirically validating our theoretical findings. |
| Researcher Affiliation | Academia | Xingyue Huang Department of Computer Science University of Oxford Oxford, UK. xingyue.huang@cs.ox.ac.uk Miguel Romero Department of Computer Science Universidad Católica de Chile & CENIA Chile mgromero@uc.cl Ismail Ilkan Ceylan Department of Computer Science University of Oxford Oxford, UK. ismail.ceylan@cs.ox.ac.uk Pablo Barceló Inst. for Math. and Comp. Eng. Universidad Católica de Chile & IMFD Chile & CENIA Chile pbarcelo@uc.cl |
| Pseudocode | No | The paper describes algorithms and update rules mathematically but does not present them in a clearly labeled pseudocode or algorithm block format. |
| Open Source Code | Yes | The code for experiments is reported in https://github.com/HxyScotthuang/CMPNN. |
| Open Datasets | Yes | We use the datasets WN18RR [31] and FB15k-237 [9], for inductive relation prediction tasks, following a standardized train-test split given in four versions [30]. We use two benchmark datasets for transductive link prediction experiments, namely WN18RR [31] and FB15k-237 [9], with the provided standardized train-test split. [...] Hetionet [15] and ogbl-biokg [16]. |
| Dataset Splits | Yes | We use the datasets WN18RR [31] and FB15k-237 [9], for inductive relation prediction tasks, following a standardized train-test split given in four versions [30]. The best checkpoint for each model instance is selected based on its performance on the validation set. |
| Hardware Specification | Yes | We ran the experiments for 20 epochs on 1 Tesla T4 GPU. All experiments are performed on one NVIDIA V100 32GB GPU. |
| Software Dependencies | No | The paper mentions software components like 'Layer Normalization' and 'ReLU' as techniques, but does not provide specific version numbers for any libraries, frameworks (e.g., PyTorch, TensorFlow), or programming languages used. |
| Experiment Setup | Yes | All models use 6 layers, each with 32 hidden dimensions. The decoder function parameterizes the probability of a fact q(u, v) as p(v | u, q) = σ(f(h(T) v|u,q)), where σ is the sigmoid function, and f is a 2-layer MLP with 64 hidden dimensions. All hyperparameter details are reported in Table 6 of Appendix C.1. |