Semi-causal decision trees

• Tipo: Artigo em Revista Científica Internacional
• Data: 2022

Autores

Nogueira, AR (Autor) Outra; Ferreira, CA (Autor) Outra; João Gama (Autor) FEP

Revista

Título: Progress in Artificial Intelligence

Vol. 11 Nº 1

Páginas: 105-119

ISSN: 2192-6352

Editora: Springer Nature

Indexação

ISI Web of Knowledge - 1 Citação

Scopus - 2 Citações

Outras Informações

ID Authenticus: P-00V-JH5

DOI: 10.1007/s13748-021-00262-2

Abstract (EN): Typically, classification algorithms use correlation analysis to make decisions. However, these decisions and the models they learn are not easily understandable for the typical user. Causal discovery is the field that studies the means to find causal relationships in observational data. Although highly interpretable, causal discovery algorithms tend to not perform so well in classification problems. This paper aims to propose a hybrid decision tree approach (SC tree) that mixes causal discovery with correlation analysis through the implementation of a custom metric to split the data in the tree's construction (Semi-causal gain ratio). In the results, the proposed methodology obtained a significant performance improvement (11.26% mean error rate) when compared to several causal baselines CDT-PS (23.67% ) and CDT-SPS (25.14%), matching closely the performance of J48 (10.20%), used as a correlation baseline, in ten binary data sets. Besides, when compared with PC in discrete data sets, the proposed approach obtained substantial improvement (16.17% against 28.07% in terms of mean error rate).

Idioma: Inglês

Tipo (Avaliação Docente): Científica

Nº de páginas: 15