Simplified Benchmark for Non ambiguous Explanations of Knowledge Graphs Link Prediction using RGCNs

Introduction

In this work, we introduce a method to benchmark explanation methods on the task of link prediction on knowledge graphs using graph neural networks (GNNs). Our approach involves constructing two specific datasets, Royalty 20k and Royalty 30k, which include an associated set of triples with each triple in the dataset. These associated triples serve as explanations for why a particular triple is considered a fact.

We conducted benchmarks using two explanation methods: GNN Explainer and ExPlaIn. The results of these benchmarks are reported for both the datasets. Importantly, for those interested in exploring these datasets further, the download links for Royalty 20k and Royalty 30k can be found below.

Dataset Details and Methodologies

Royalty 20k and Royalty 30k

• Royalty 20k: Consists of 20,000 triples.
• Royalty 30k: Consists of 30,000 triples.
• Each dataset is annotated with related triples explaining the factuality of each asserted triple.

Explanation Methods Benchmarked

1. GNN Explainer: A method dedicated to providing explanations for models built upon graph neural networks.
2. ExPlaIn: A framework focused on generating explanations for GNN-based models, providing insight into the decision-making process of the network.

Results

• The benchmarking results for GNN Explainer and ExPlaIn are detailed in the accompanying publication.

Access to Datasets

• Download links for both Royalty 20k and Royalty 30k datasets are provided for further research and exploration.

Conclusion

Through our research, we offer a simplified and structured method to benchmark explanation techniques for link prediction in knowledge graphs, particularly using relational graph convolutional networks (RGCNs). This enables a clearer understanding of how certain knowledge is deduced by the networks, paving the way for more interpretable AI models.

Keywords

• Knowledge Graphs
• Link Prediction
• Graph Neural Networks (GNNs)
• GNN Explainer
• ExPlaIn
• Relational Graph Convolutional Networks (RGCNs)
• Royalty 20k
• Royalty 30k

FAQ

Q1: What is the purpose of the Royalty 20k and Royalty 30k datasets? A1: These datasets are designed to assist in benchmark testing of explanation methods for link prediction tasks in knowledge graphs using GNNs.

Q2: What explanation methods are benchmarked in this study? A2: The study benchmarks two methods: GNN Explainer and ExPlaIn.

Q3: For what applications can these datasets be particularly useful? A3: These datasets are particularly useful for researchers focusing on interpretability in AI, especially in applications involving knowledge graphs and GNNs.

Q4: Where can the datasets be accessed? A4: The datasets can be downloaded via the links provided in the publication.

Q5: Why is it important to benchmark explanation methods for knowledge graphs? A5: Benchmarking allows for the assessment of how well explanation methods can provide transparent and interpretable reasons behind the decisions made by GNN models, which is crucial for trust and reliability in AI applications.