Graph Laplacian Based Ontology Regularization Distance Framework for Ontology Similarity Measuring and Ontology Mapping

Linli Zhu, Yu Pan, Mohammad Reza Farahani, Wei GaoJournal of Computational Mathematica 2017 · DOI

TL;DR

This paper introduces a novel framework for measuring ontology similarity and performing ontology mapping by learning an optimal distance metric. It leverages graph Laplacian techniques to incorporate unlabeled ontology data, enhancing the robustness and accuracy of the similarity calculations. The proposed method demonstrates improved performance in experiments on plant and humanoid robotics ontologies compared to existing approaches.

What Problem Does This Paper Address?

Ontology engineering applications, such as information retrieval and semantic computation, heavily rely on accurately measuring the similarity between concepts within ontologies and mapping concepts between different ontologies. Existing methods often represent concepts as high-dimensional vectors and compute similarity based on these vectors. However, effectively capturing the semantic relationships and leveraging all available information, including unlabeled data, remains a challenge. This paper addresses the need for a more robust and accurate ontology similarity and mapping framework by learning an optimal distance metric that utilizes both labeled (similar/dissimilar pairs) and unlabeled ontology data.

Methods

The core of the proposed method is a regularization framework for learning an ontology distance metric. This involves defining an objective function that minimizes the distance between similar concept vectors and maximizes the distance between dissimilar concept vectors. The authors introduce a graph Laplacian regularizer to effectively incorporate unlabeled ontology data. This regularizer utilizes the structure of the ontology graph, represented by a weight matrix derived from nearest neighbor relationships (based on Euclidean distance initially), and a Laplacian matrix. The framework aims to learn a symmetric positive semi-definite (PSD) distance matrix W that defines the geometric distance between concept vectors. The optimization problem is formulated to find the optimal W by minimizing a combination of loss functions on labeled data and the graph Laplacian regularizer.

Key aspects of the method include:

  • Geometric Distance: Similarity is measured by the geometric distance between concept vectors in a p-dimensional space.
  • Regularization Framework: A min-max approach is used to learn the distance metric, minimizing distances for similar pairs and maximizing for dissimilar pairs.
  • Graph Laplacian Regularization: This novel component leverages unlabeled ontology data by constructing a weight matrix based on nearest neighbors and then using the graph Laplacian to regularize the learning process. This helps to prevent overfitting and improve generalization.
  • Optimization: The problem is framed as a convex optimization problem, with specific attention paid to avoiding trivial solutions (e.g., W=0) by incorporating a log-determinant constraint.

Datasets

The paper does not introduce new datasets. Instead, it uses existing ontologies for evaluation:

  • "PO" ontology: Used for the ontology similarity measuring experiment. This ontology can be found at http://www.plantontology.org.
  • "Humanoid Robotics" ontologies (O2 and O3): Used for the ontology mapping experiment. The structure of these ontologies is referenced from Gao Zhu [24].

The experiments utilize expert-defined sets of similar (S) and dissimilar (D) vertex pairs for training and evaluation.

Key Insights and Ideas

  • Unlabeled Data is Crucial: The paper highlights that unlabeled ontology data, when effectively incorporated through graph Laplacian techniques, significantly improves the performance of ontology similarity and mapping tasks. This is a key differentiator from methods that solely rely on labeled pairs.
  • Learned Metric Outperforms Fixed Metrics: By learning an optimal distance metric W, the framework can adapt to the specific characteristics of the ontology data, leading to more accurate similarity measurements than using a fixed metric like Euclidean distance.
  • Robustness through Regularization: The regularization framework, particularly with the graph Laplacian component, helps to make the learned distance metric more robust and less prone to overfitting, which is essential for real-world applications.
  • Unified Framework for Similarity and Mapping: The proposed approach provides a unified framework that can be applied to both ontology similarity measurement and ontology mapping, demonstrating its versatility.

Key Results

The paper presents experimental results comparing their proposed algorithm against three existing methods ([11], [12], [13]) on two distinct domains:

  • Plant Ontology Similarity: In the plant ontology similarity experiment, the proposed algorithm consistently achieved higher precision ratios (P@3, P@5, P@10) compared to the other algorithms. For instance, at P@10, the proposed method achieved a precision ratio of 0.6207, while the best performing baseline reached 0.6170. The differences became more pronounced as the number of considered neighbors (N) increased.
  • Humanoid Robotics Ontology Mapping: Similarly, for humanoid robotics ontology mapping, the proposed algorithm outperformed the baselines across P@1, P@3, and P@5 precision ratios. For example, at P@5, the proposed method achieved 0.6778, significantly higher than the best baseline at 0.5385.

Limitations and Future Work

  • Computational Complexity: The paper mentions that the complexity of the ontology learning algorithm is O(n²), where n is the size of the ontology data. While this is stated, the practical implications of this complexity for very large ontologies are not deeply explored.
  • Initial Euclidean Distance: The weight matrix w for incorporating unlabeled data is initially determined based on standard Euclidean distance. The paper does not explore alternative initial distance metrics that might be more suitable for certain ontologies.
  • Future Work: The paper implicitly suggests that further exploration of regularization techniques and optimization strategies could lead to even better performance. The successful application in two distinct domains also opens avenues for testing on a wider range of ontologies and applications.

Summary updated

Abstract

The core of a large number of ontology engineering applications is,similarity measuring, and the essence of ontology mapping and its applications are also similarity computation. There are a great variety of learning techniques are designed for ontology similarity calculating in different engineering circumstances. In these learning settings, all the semantic information of concept is enclosed in the p dimensional vector. In this paper, the ontology similarity measuring is considered via geometric distance calculating between the vectors of ontology vertex corresponding to. We use regularization framework to solute the optimal ontology distance function and apply graph Laplacian to make full use of the unlabeled ontology information. Finally, two simulation experiments imply that our graph Laplacian based ontology regularization distance framework works well in plant and humanoid robotics ontology applications in which higher precision ratio results are obtained from our algorithm than previous ontology learning tricks.
BibTeX
@article{zhu2017graph,
  author = {Linli Zhu and Yu Pan and Mohammad Reza Farahani and Wei Gao},
  title = {{Graph Laplacian Based Ontology Regularization Distance Framework for Ontology Similarity Measuring and Ontology Mapping}},
  journal = {Journal of Computational Mathematica},
  year = {2017},
  doi = {10.26524/cm6},
}
Figure