Rule-guided GNNs for Explainable KG Reasoning

This is the official code of the paper Rule-guided GNNs for Explainable Knowledge Graph Reasoning (AAAI'25).

Overview

We proposed a new family of GNNs whose training can be guided by rules. This is achieved through a connection between the parameters of our GNN with various forms of rules. Specifically, our approach allows

• effective rule injection to guide the GNN training,
• efficient rule extraction by directly "reading off" the GNN parameters,
• a comparison between the two sets of rules to analyse the effectiveness of our rule guidance, and
• the extracted rules that are both sound and plausible.

Here are examples to demonstrate the connection between the rules and the GNN parameters. Consider a rule $p_i(x,y) \leftarrow p_j(x,y)$ with confidence $\alpha$, and a pair of entities $a,b$. If the fact $p_j(a,b)$ holds, then $h_{a,b}^{(0)}[j]=1$ according to our KG encoding. From our rule encoding, $A^{(1)}[i][j]=\alpha$. From the message passing equation, $h_{a,b}^{(1)}[i] \ge \big(A^{(1)}h^{(0)}_{a,b}\big)[i] \ge \alpha$, which means the plausibility of the fact $p_i(a,b)$ is greater than that of $p_j(a,b)$.

Installation

• python: 3.7.7
• pytorch: 1.12.1
• torch_geometric: 2.3.1

Directory Structure

The following basic directory structure is required to run our implementation:

.
├── configs
├── data
    └── GraIL-BM_fb237_v1
        └── train-labeled-typed.txt
        └── valid-labeled-typed.txt
        └── test
            └── test_graph_w_types.txt
            └── test0.txt
            └── test0_with_truth_values.txt
            └── ...
            └── test9.txt
            └── test9_with_truth_values.txt
    └── ...
├── experiments
├── predicates
├── rules

Reproduction

We provided the configuration files in the configs folder to reproduce the results in the paper.

Training

To train our model, run the script ./train.py. The --config argument is the file path of the configurations. The --gpu argument indicates which gpu to use, which is optional. As an example, here is training command for the benchmark GraIL-BM_fb237_v1.

python train.py --config GraIL-BM_fb237_v1-train.yaml 

This will store the learned model in experiments/${experiment_name}/models/model.pt and save a snapshot model in experiments/${experiment_name}/models/e{epoch}.pt once every 100 epochs.

Applying a model

To apply a model to a dataset, run the script ./evaluate.py. Similarly, the --config argument is the file path of the configurations for test, and the --gpu argument indicates which gpu to use.

As an example, here is the command for applying the model trained for the benchmark GraIL-BM_fb237_v1.

python evaluate.py --config GraIL-BM_fb237_v1-eval.yaml 

Rule extraction

To extract the rules captured by our model and compare the rules with those from the baselines, run the script ./extract_rules.py.

 python evaluate_rules.py

This will output the number of extracted rules, the percentage of high-quality rules of each method for each benchmark.

Citation

If you find the code useful in your research, please cite the following paper.

@inproceedings{wang2025rule,
  title={Rule-guided GNNs for Explainable Knowledge Graph Reasoning},
  author={Wang, Zhe and Ma, Suxue and Wang, Kewen and Zhuang Zhiqiang},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  year={2025}
}