IndKGC: Inductive Knowledge Graph Completion with GNNs and Rules:An Analysis

This is the official implemetation of the paper Inductive Knowledge Graph Completion with GNNs and Rules: An Analysis. This paper exloits RGCN and CompGCN as GNNs and AnyBURL as rule learner. The CompGCN convolution layers are used from the official implementation of CompGCN paper. Thanks to the authors of CompGCN for sharing the updated code-base.

Authors: [Akash Anil, Víctor Gutiérrez-Basulto, Yazmín Ibáñez-García, Steven Schockaert]

Overview

IndKGC presents an analysis over Knowledge Graph Completion in inductive setting where Training and Testing knowledge graphs are disjoint in terms of entities. This setting forces the models to predict the future connections or links only using the graph strcuture (relation, induced graphs, etc.) not the entities. Looking at the inductive scenario, rule-based link prediction seems to be a natural choice. However, the state-of-the-art methods are only based on the GNN frameworks. This paper first revisits the rule-based method namely, AnyBURL and evaluates the link prediction in inductive setting. We find that the rule-based methods are unfairly disadvataged by the current evaluation setting. Further, rule-based methods are limited to only the predicted rules. Thus, we propose a hybrid strategy to use rules and gnn in unison. We propose a GNN framework based on CompGCN/R-GCN GNNs which is trained over the rule path subgraphs.This hybrid strategy i.e., GNN + Rule boosts the performance of link prediction for AnyBURL as well as preserve the interpretability.

Dependencies Installation

You can install the required dependencies using either conda or pip. IndKGC shall work with Python >= 3.7 and PyTorch >= 1.8.0.

Using Conda

conda install pytorch=1.8.0 cudatoolkit=11.1 pyg -c pytorch -c pyg -c conda-forge

Using Pip

pip install torch==1.8.0+cu111 -f https://download.pytorch.org/whl/torch_stable.html
pip install torch-scatter==2.0.8 torch-sparse==0.6.12 torch-geometric -f https://data.pyg.org/whl/torch-1.8.0+cu111.html

Dataset

IndKGC is evaluated over three inductive benchmark datasets namely, FB15k-237, WN18RR, and NELL-995 provided by the seminal paper Inductive Relation Prediction by Subgraph Reasoning. All of these datasets have been divided into four versions v1, v2, v3, and v4. The training and validation graphs are extracted from data_version (e.g., fb237_v1) and the test datasets such as fact graph (train_ind) and test_ind are extracted from data_version_ind (e.g., fb237_v1_ind) similar to NBFNet.

Pre-requisite Information

1. AnyBURL

IndKGC requires AnyBURL rules and predictions over train, valid, and test_ind for each dataset used.

2. NBFNet Rank & Test Score

IndKGC require NBFNet ranks and model scores for test data. The NBFNet code should be modified to store the ranks and test data scores in the same order of the test data considered. For example, IndKGC orders the test triplet first and then the corresponding inverse triplets (first <h,r,t> followed by <t, inv_r, h>) from train_ind. Thus, to maintain the order, NBFNet code should be modified to store the ranking and scores in the similar fashion.

Reproduction

The main results in the IndKGC are obtained using the following steps:
a.0. Get the rule path instantiations for CompGCN/RGCN for the triplets of train, valid, and test_ind datasets.

   python3 script/get_rule_paths.py -d data/fb15k237_v1 -r anyburl-22/fb15k237_v1

a.1. Get noisy-or confience for Test data

   python3 script/get_noisy_or.py -d data/fb15k237_v1 -r anyburl-22/fb15k237_v1

b. Train and Test CompGCN/RGCN/Noisy-OR model over the rule paths.

   python3 script/run.py -d data/fb15k237_v1

   python3 script/run_rgcn.py -d data/fb15k237_v1

   python3 script/run_noisy.py -d data/fb15k237_v1

c. The above execution will evaluate Noisy-OR and train and evaluate CompGCN/R-GCN, CompGCN/R-GCN + NBFNet, and NBFNet + NBFNet models. To get evaluation for AnyBURL and AnyBURL + NBFNet execute the below command

   python3 script/eval_anyburl.py --data_dir data/fb15k237_v1 -r anyburl-22/fb15k237_v1

d. The NBFNet evaluation can be obtained by running NBFNet.

If you make use of this code in your work, please cite the following paper:

@misc{anil2023inductive,
      title={Inductive Knowledge Graph Completion with GNNs and Rules: An Analysis}, 
      author={Akash Anil and Víctor Gutiérrez-Basulto and Yazmín Ibañéz-García and Steven Schockaert},
      year={2023},
      eprint={2308.07942},
      archivePrefix={arXiv},
      primaryClass={cs.AI}
}

anilakash/IndKGC