Abstract
AutoRDF2GML is a framework designed to bridge the "semantic gap" between RDF Knowledge Graphs and Graph Machine Learning (GML). By automating the extraction of both literal content and graph topology into machine-learning-ready formats, the framework enables researchers to leverage Semantic Web data without extensive domain expertise in SPARQL or RDF internals. This paper introduces the framework's architecture and four new benchmark datasets that offer richer feature sets than existing industry standards.
Introduction
The Semantic Web provides a vast repository of structured data in the Resource Description Framework (RDF). However, utilizing this data for Graph Neural Networks (GNNs) remains challenging due to the heterogeneity and sparsity of RDF data. GML models typically require data in the form of tensors and feature matrices, whereas RDF data is often composed of complex literals (strings, dates, numbers) and sparse relational links. Currently, researchers must manually engineer features via complex SPARQL queries, a process that is both error-prone and time-consuming. AutoRDF2GML addresses this by automating the "propositionalization" process, turning relational data into feature vectors.
The AutoRDF2GML Framework
The framework is designed to handle the heavy lifting of data conversion through a semi-automated pipeline. Users provide an RDF dump and a simple configuration file, and the system generates a GML-ready dataset.
Core Methodology
- Automated Feature Selection: The system automatically identifies useful RDF properties, filtering out those that are too sparse or contain unique identifiers that might lead to overfitting.
- Content-Based Feature Extraction: It transforms various literal types into vectors. For instance, it employs SciBERT to encode natural language descriptions and uses one-hot or label encoding for categorical data.
- Topology-Based Feature Extraction: To capture the structural "shape" of the graph, the framework uses Knowledge Graph Embedding (KGE) techniques like TransE, converting entity relationships into dense vectors.
- Handling Complex Relations: Unlike many existing tools, AutoRDF2GML can model n-ary relations (where links have their own attributes) and multi-hop relations, preserving the rich metadata often lost during conversion.
New Benchmarks for GML
A significant contribution of this work is the introduction of four new benchmark datasets created using the framework. These datasets are designed to be richer in features than standard GML benchmarks, specifically providing "Literals\NLD" (non-textual literals) and "Topology" features across almost all node types.
- SOA-SW (SemOpenAlex-SemanticWeb): An academic publication dataset with ~171k nodes and 1.8M edges.
- LPWC (Linked Papers With Code): Links papers to specific tasks, methods, and datasets, featuring ~391k nodes.
- AIFB: A classic reasoning dataset enriched with automated features.
- LinkedMDB: An RDF-based version of the Internet Movie Database.
Key Insights and Results
By providing standardized, automatically generated features for both content and topology, AutoRDF2GML allows for more rigorous model comparison. It removes the ambiguity of whether a GNN's performance is due to its architecture or clever manual feature engineering. Initial evaluations on recommendation tasks using these enriched features showed measurable improvements in F1 scores compared to using structural data alone.
Conclusion and Future Work
AutoRDF2GML lowers the barrier to entry for the machine learning community to utilize Knowledge Graphs. While currently semi-automatic, future iterations aim to enhance scalability across the Linked Open Data cloud and integrate OWL reasoning. Incorporating ontological rules would allow the framework to automatically merge equivalent entities or infer missing links before the machine learning stage, further narrowing the semantic gap.
Summary updated