A carregar...
Publicação
Identification of Logical Inference Patterns in Knowledge Networks
| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 2.96 MB | Adobe PDF |
Autores
Orientador(es)
Resumo(s)
Understanding how variables influence one another is a central challenge in data analysis, as it requires moving beyond associations to uncover the direction of relationships. This directional structure is essential for interpreting data, reasoning about causality, and building models that generalize to new situations. Directed Acyclic Graphs (DAGs) provide a natural representation of these relationships, but existing structure learning methods are often complex, computationally intensive, and fail to fully recover the true graph as data complexity increases. In this thesis, we study this problem using benchmark datasets with predefined structures, generating data from known graphs to apply and evaluate multiple structure learning algorithms. This allows us to identify the strengths and weaknesses of current methods and assess their practical performance. To complement existing approaches, we develop a logic-based, implication-driven method that evaluates pairs of variables locally using confusion matrices, producing scores that indicate potential directions of influence. This pairwise evaluation is combined with a correlation metric to detect associations, and the method operates on binary data to simplify calculations and improve interpretability. Each pairwise decision can be fully traced and understood, offering transparency that is often missing in traditional structure learning methods, which rely on complex global optimization and probabilistic assumptions. While results show that the method performs less accurately on small benchmark datasets, its performance becomes comparable to established approaches on larger datasets, and its computational cost remains practical due to the simplicity of the calculations. This approach is explainable and systematic, and it can serve as a building block for further work in structure learning, providing a simple and efficient method that performs reliably under favorable conditions in an area central to understanding data and making inferences, where further research is still needed to improve existing techniques.
Descrição
Tese de mestrado, Engenharia Informática, 2025, Universidade de Lisboa, Faculdade de Ciências
Palavras-chave
Structure Learning Directed Acyclic Graphs (DAGs) Scalability Conditional Probability Causality