Quality of an Ontology as a Dynamic Optimisation Problem

Michael Cochez, Vagan Y. TerziyanICTERI 2012 · OpenReview

Introduction

The traditional vision of the Semantic Web relies on a "Holy Grail" of universal agreement: a world where all data is machine-processable because it adheres to shared, static ontologies. However, this approach faces significant practical hurdles in real-world deployment. This paper proposes a paradigm shift, reframing ontology selection and maintenance not as a search for a single, universal truth, but as a dynamic, multi-objective optimization problem. By factoring in the specific context of an application—such as hardware constraints or geographic location—systems can automatically adapt or approximate the "best" ontology for a given task.

The Limitations of Static Ontologies

The authors identify two primary obstacles to the widespread adoption of rigid, shared ontologies:

  1. The Agreement Bottleneck: Reaching a consensus on a single cataloging scheme across diverse groups is notoriously difficult. In many real-world scenarios, ontologies are either incomplete, partially known, or entirely missing.
  2. Human Subjectivity: Rigid schemas struggle to capture "fuzzy" human-oriented concepts like "happiness" or "beauty." These subjective notions do not fit neatly into the binary, static classifications typical of traditional logic.

Currently, when an ontology is missing or ill-suited, systems often reach a standstill. This research explores how a system might instead "approximate" the necessary parts of an ontology based on the data it encounters and the specific environment (context) in which it operates.

Methodology: A Mathematical Framework

As a theoretical work-in-progress, this paper establishes a mathematical foundation for ontology quality rather than a specific algorithm. The proposed framework rests on three pillars:

Multi-objective Optimization

"Quality" is not a monolithic metric. An ontology may excel in coverage (detail) while failing in coupling (complexity). The framework treats these competing features as objectives in an optimization problem, seeking a balance that serves the specific needs of the system.

Dynamic Context-Sensitivity

The "optimal" ontology is a moving target. For instance, a mobile device's requirements change based on its environment. A system operating near a coastline might prioritize concepts like "shores" and "harbours," but as the user moves into a mountainous region, it should swap these for "peaks" and "valleys" to conserve memory and processing power.

Fuzzy Logic Integration

To bridge the gap between machine logic and human subjectivity, the authors suggest using fuzzy ontologies. In these models, class membership is represented as a grade between 0 and 1, allowing for more nuanced, human-like data processing and representation of ill-defined concepts.

Key Insights: Features vs. Quality

A critical distinction made in the paper is between features and quality. Features are intrinsic properties of an ontology, such as cohesion, depth, or coupling. Quality, however, is the measure of how well those features serve a specific context. A small, simple ontology is considered "higher quality" than a massive, comprehensive one if the target device has limited storage and processing power.

This perspective links ontology theory to "Executable Reality" and self-managed systems. In Mobile Mixed Reality settings, for example, a system must dynamically prune or expand its active ontology to remain efficient as the user's context evolves. This implies that the border between schema and instance data is blurry; the data itself should help shape the ontology over time.

Formalizing the Optimization Problem

The paper’s primary contribution is a formal definition of the context-dependent dynamic multi-objective optimization problem. It defines a solution ($sol$) as a function mapping a set of contexts ($C$) to a set of optimal ontologies ($O$):

$$sol: C \to O$$

The authors acknowledge that in real-world systems, this function can likely only be known by approximation. Future contexts are often unknown, and systems cannot always react instantaneously to every complex environmental change.

Limitations and Future Work

While the framework provides a robust theoretical base, several challenges remain for future research:

  • Implementation: The paper does not yet specify concrete ontology notations (such as specific OWL2 profiles) or specific search algorithms to find the optimal ontology.
  • Defining Context: A concrete mathematical representation of "context"—including how to represent user intent or location within the optimization function—remains an open question.
  • Computational Overhead: The cost of frequently replacing or updating an ontology (consistency checking and data re-processing) must be balanced against the benefits of a better-fitting model.
  • System Alignment: If independent systems evolve their own "optimal" ontologies, ensuring they can still communicate (ontology alignment) becomes a critical hurdle.

Summary updated

BibTeX
@inproceedings{cochez2012quality,
  author = {Michael Cochez and Vagan Y. Terziyan},
  title = {{Quality of an Ontology as a Dynamic Optimisation Problem}},
  booktitle = {ICTERI},
  year = {2012},
  url = {https://openreview.net/forum?id=V9CtdX7dnQo},
}
Figure