Título: Causality -Driven Feature Engineering for Enhanced Telecommunications Churn Prediction
Autores: Gustavo Viegas, Marcus Mendes & Fabrício Aguiar Silva
Resumo: Customer churn prediction is critical for telecommunication companies, but traditional machine learning models often struggle with the minority churn class and rely on correlational features. A possible solution to tackle this problem is to use causal-based feature engineering, which considers the causal relationship between the features and the churn decision. This study employs a causality framework to generate causal, tailored feature sets that enhance predictive accuracy, particularly for the minority churn class. Our findings indicate that feature sets engineered with causal insights significantly improve the discovery of underlying causal structures. In predictive tasks, these causally informed feature sets, especially when combined with oversampling, consistently outperformed baseline configurations in both test and cross-validation setups. These results demonstrate that integrating causal discovery into feature engineering is a potent strategy for developing more accurate, stable, and insightful churn prediction models.
Palavras-chave: Causality; causal discovery; customer churn prediction; machine learning; causal feature engineering.
Páginas: 8
Código DOI: 10.21528/CBIC2025-1175578
Artigo em PDF: CBIC_2025_paper1175578.pdf
Arquivo BibTeX:
CBIC_2025_1175578.bib