Rule Hierarchy Framework

Overview

This repository contains the codebase and datasets as described in paper "Building Rule Hierarchies for Efficient Logical Rule Learning from Knowledge Graphs". We introduce the Rule Hierarchy Framework (RHF) that efficiently organizes the rules mined by walk-based (bottom-up) logical rule learners into a proper rule hierarchy. In this repository, we augment the GPFL system with the adaptation of RHF and the extensions of two hierarchical pruning methods (HPMs) that utilize the generated rule hierarchy to prune irrelevant rules. The Figure above shows an incomplete rule hierarchy where rules are linked by subsumption relations.

Requirements

• Java >= 1.8
• Gradle >= 5.6.4

Getting Started

To verify the system, run:

gradle run --args="-c data/UWCSE/config.json -r"

which executes:

• Rule Learning: learn rules for target predicates. The learned rules in this example can be found at data/UWCSE/i3c3/rules.txt.
• Rule Application: Apply the learned rules to infer new facts. The inferred facts can be found at data/UWCSE/i3c3/predictions.txt, and in file data/UWCSE/i3c3/verifications.txt you can find the explanations for both inferred and known facts.
• System Evaluation: Evaluate the learned rules in filtered hits@n and MRR. Detailed report can be found at data/UWCSE/i3c3/eval_log.txt.

To turn on/off the prior and post pruning and tune the prior threshold, please change the configuration file data/UWCSE/config.json by setting:

• use_prior_prune: true/false. Enable the execution of the preparations that are related to the prior pruning.
• prior_th: set the prior threshold.
• use_post_prune: true/false. Enable the post pruning.

For details about other settings, please refer to GPFL Codebase.

Reproduce Experiment Results

Please download the datasets and unzip the compressed files into the data folder. We recommend running the experiments with at least 8 CPU cores and 64GB RAM for completion in reasonable time.

Over datasets {FB15K-237-LV, WN18RR-LV, NELL995-LV, OpenBioLink}, one can evaluate the effectiveness of the prior and post pruning by tuning prior_th and use_post_prune and keep all other settings unchanged in their individual configuration file, and run:

gradle run --args="-c data/[dataset]/config.json -r"