- Titel
- Complementary Extensions of Concept Learning in Description Logics
- Zitierfähige Url:
- https://nbn-resolving.org/urn:nbn:de:bsz:15-qucosa2-994729
- Datum der Einreichung
- 05.02.2025
- Datum der Verteidigung
- 24.09.2025
- Abstract (EN)
- Instead of a sheer amount of data, having information with explicit semantics is a key enabler for turning data into knowledge. Data provided in semantically explicit formats allows machines to calculate insights from it. A technique that leverages the explicit semantics and automated reasoning capabilities is Concept Learning (CL). The broad idea of CL is to learn concept descriptions from data. It has proven successful in many real-world use cases and is advantageous over black-box systems. In many application fields, a learned classifier that can provide further insights, besides mere classification capabilities, is favorable. However, despite its solid formal foundation, we see unsolved challenges in Concept Learning. A first challenge (Ch1) concerns the lack of solutions for a systematic evaluation of CL algorithms. Another challenge (Ch2) is the size of the search space, especially in application domains that are modeled in great detail, and, hence, have a very large number of defined concepts. In many existing CL algorithms, this causes a ‘combinatorial explosion’ regarding the generated candidate concepts. This circumstance requires new exploration strategies to cover the search space more efficiently. Finally, we see a rise in the amount of spatial data. However, the knowledge representation formalisms underlying Concept Learning usually do not cover the extraction of spatial relations. We see the integration of these implicit relations into CL as a further challenge (Ch3). To address Ch1, we propose a framework for the automatic evaluation of different CL algorithms. We perform a systematic literature review to find a diverse and established set of CL scenarios. We implement a free software system called SML-Bench to run a suite of freely configurable benchmark runs and get one unified evaluation result. Regarding Ch2, we found that it is worth considering mechanisms for balancing the explorative and exploitative behavior of CL algorithms, especially in cases involving more extensive background knowledge bases. While more explorative algorithms typically cover a wider area of the search space and thus generate a more diverse set of hypotheses, algorithms dominated by exploitative behavior focus on areas where the greatest improvement is achieved. However, in their extremes, a fully explorative, i.e., random, CL algorithm would be inefficient, and greedy, purely improvement-driven algorithms would be prone to getting stuck in local optima. To better fine-tune the search space exploration behaviors of CL algorithms, we reviewed the literature on metaheuristics. As a first in-depth study, we focus on a CL approach based on the Simulated Annealing meta-heuristic and demonstrate performance improvements that surpass the existing state-of-the-art. To fully utilize spatial information in Concept Learning and address Ch3, we focus on a set of 20 spatial relations, define them formally, demonstrate how to automatically infer these relations from polygon data, and explain how to integrate them into the CL formalism. Furthermore, we discuss their computational complexities. We apply this extension in different experiments to show its effectiveness.
- Freie Schlagwörter (EN)
- Concept Learning, Description Logics
- Klassifikation (DDC)
- 500
- Den akademischen Grad verleihende / prüfende Institution
- Universität Leipzig, Leipzig
- Version / Begutachtungsstatus
- angenommene Version / Postprint / Autorenversion
- URN Qucosa
- urn:nbn:de:bsz:15-qucosa2-994729
- Veröffentlichungsdatum Qucosa
- 06.10.2025
- Dokumenttyp
- Dissertation
- Sprache des Dokumentes
- Englisch
- Lizenz / Rechtehinweis
CC BY-SA 4.0- Inhaltsverzeichnis
1 Introduction 1.1 Motivation 1.2 Research Questions 1.3 Overview of the Thesis 2 Preliminaries 2.1 Knowledge Representation and Reasoning 2.1.1 Foundations 2.1.2 Ontology and Ontologies 2.1.3 Automated Reasoning 2.2 First-Order Logic 2.3 Horn Clause Logic and Logic Programming 2.3.1 Logic Programming 2.3.2 Prolog 2.4 Description Logics 2.4.1 Relation Between Description Logics and Logic Programming 2.5 Semantic Web Technologies 2.6 Machine Learning 2.7 Inductive Logic Programming 2.8 Concept Learning 3 Related Work 3.1 Concept Learning 3.1.1 Inductive Learning 3.1.2 Inductive Learning in Logic Programming (ILP) 3.1.3 Inductive Learning in Description Logics and OWL (Concept Learning) 3.2 Benchmarking Concept Learning 3.3 Metaheuristics for Concept Learning 3.4 Spatial Concept Learning 4 The DL-Learner Framework 4.1 Learning Problems 4.2 Overview of the Framework 4.3 Learning Algorithms 4.3.1 Refinement Operator-based Algorithms 4.3.2 OWL Schema Learning Algorithms 4.3.3 Other Algorithms 4.4 Implementation 4.5 Contributions 5 Benchmarking Concept Learning 5.1 Challenges of Structured Machine Learning 5.2 Benchmarking Structured Machine Learning Algorithms 5.3 Datasets 5.4 SML-Bench Framework 5.4.1 Architecture 5.4.2 Benchmark Scenarios 5.4.3 Benchmarked Tools 5.4.4 Benchmark Settings 5.4.5 Available Learning Systems 5.5 Evaluation 5.6 Discussion 6 Extended Metaheuristics for Concept Learning 6.1 Approach 6.1.1 Refinement-based Concept Learning in Description Logics as a Hill Climbing Approach 6.1.2 Simulated Annealing 6.1.3 Simulated Annealing in Concept Learning 6.2 Empirical Evaluation 6.2.1 Evaluation Setup 6.2.2 Results 6.3 Discussion 7 Spatial Concept Learning 7.1 A Calculus for Polygon Data 7.2 Spatial Inference on Geospatial Polygon Data in Description Logics 7.3 Spatial Concept Learning 7.4 Implementation 7.5 Evaluation 7.5.1 Spatial Inference 7.5.2 Spatial Concept Learning 8 Conclusions 8.1 The DL-Learner Framework 8.2 Benchmarking Concept Learning 8.3 Extended Metaheuristics for Concept Learning 8.4 Spatial Concept Learning A Curriculum Vitae Bibliography Index