Advanced property inference in RDF graphs using approximate and generalized bonds

RDF
Semantic web
FCA
Bonds
Approximate reasoning
Algorithm
Publication idea

This paper extends our prior work on RDF property inference by applying two new theories: approximate bonds and intermediate-quantifier bonds. This allows our FCA-based method to handle noisy data more robustly and gives users fine-grained control over the inference process. We demonstrate this provides a more flexible and powerful tool for knowledge graph completion.

Author

Domingo López Rodríguez, Marek Reformat, Ondrej Krídlo, Manuel Ojeda Aciego

Published

1 December 2025

Keywords

Property inference, Knowledge graph completion, Approximate bonds, Intermediate quantifiers, RDF, DBpedia, Link prediction

Property inference in RDF graphs—the task of deducing new triples (subject, predicate, object) from existing data—is crucial for knowledge graph completion. Our prior work established a method for this using FCA bonds based on rigorous (universal) and benevolent (existential) quantifiers. While effective, this approach is limited by the crisp nature of standard bonds and the inflexibility of the quantifiers. This paper significantly extends this framework by applying our newly developed theories of approximate bonds and bonds generated by intermediate quantifiers to the RDF inference problem. By using approximate bonds, our method can infer properties even in the presence of noisy or incomplete RDF data. By using intermediate quantifiers, an expert user can fine-tune the inference process, specifying the desired level of evidence required to deduce a new relation. We demonstrate on a large-scale RDF benchmark that these advanced bond-based methods provide a more robust and flexible tool for semantic inference.

Introduction

RDF graphs are the backbone of the Semantic Web. A key challenge is inferring new knowledge, i.e., new edges (properties) in the graph. Our previous work, currently in draft, introduced a novel method using FCA bonds: an RDF graph is decomposed into bipartite graphs (formal contexts), and bonds are computed between them based on external information to infer new triples for a target property.

However, that work was limited to the two extreme bond types. This paper serves as the direct, applied follow-up, leveraging two of our new theoretical constructs:

  1. Approximate bonds: To handle the inherent noise and incompleteness of real-world RDF data.
  2. Bonds from intermediate quantifiers: To allow a user to express more nuanced inference rules, moving beyond “all” or “at least one.”

The goal is to show that these more advanced bond theories translate directly into a more powerful and practical RDF inference engine.

Methodology

The core methodology follows our previous work, but with the bond computation step replaced. 1. An expert user identifies a target property (relation) to infer. 2. The user identifies the source bipartite graphs (contexts) from the RDF data. 3. The user provides external linking information. 4. Novelty: The user now also specifies either a tolerance threshold (for an approximate bond) or an intermediate quantifier (e.g., “at least 50%”). 5. Our system computes the corresponding advanced bond. The triples in this bond represent the inferred properties.

Experimental validation

We will use a standard knowledge graph completion benchmark (e.g., a subset of DBpedia or Freebase). We will compare the inference results (precision, recall, F1-score) of our new methods against: * Our original rigorous/benevolent bond method. * Standard statistical link prediction methods (e..g., TransE, ComplEx).

The goal is to show that our method is competitive in accuracy while being more flexible and interpretable.

Work plan

  • Months 1-3: Implement the computation of approximate bonds and intermediate-quantifier bonds within the RDF inference pipeline.
  • Months 4-7: Run experiments on the chosen RDF benchmark.
  • Months 8-10: Analyze the results, focusing on cases where the new methods succeed and the old ones fail (e.g., in noisy data).
  • Months 11-12: Write the manuscript, targeting a top journal in the Semantic Web or AI field.

Potential target journals

  1. Knowledge-Based Systems (Q1): An excellent target, as the work is a direct application of a novel knowledge representation technique to a practical problem. Your draft is already targeting this journal.
  2. Journal of Web Semantics (Q1): The premier journal for research related to the Semantic Web and RDF.
  3. Expert Systems with Applications (Q1): A strong candidate for its focus on practical AI systems.

MVA strategy

This is a natural application paper that builds on prior theoretical work.

  • Paper 1 (The MVA):
    • Scope: A single, strong paper that introduces the application of both approximate and intermediate-quantifier bonds to RDF inference. It should clearly show the added value of each technique through well-designed experiments.
    • Goal: To demonstrate the practical power of the new bond theories and establish bond-based inference as a flexible, state-of-the-art method for RDF knowledge graph completion.
    • Target venue: Knowledge-Based Systems.