31Jul2026
09:00 Master's Defense room 85 of IC2
Topic on
CAUSE-GNN: Causally constructed similarity graph-based deep learning applied to churn prediction.
Student
Mariana Aparecida Ferreira
Advisor / Teacher
Marcelo da Silva Reis - Co-advisor: Julio Cesar dos Reis
Brief summary
Churn prediction is a highly relevant problem for companies in various sectors seeking sustainable financial growth, especially in the financial services and telecommunications areas. In these domains, relationships between customers and product consumption tend to be highly complex, resulting in large volumes of data and dependency structures that often hinder modeling using traditional methods. Classical machine learning approaches, while effective in predominantly tabular scenarios, tend to fail to capture structural dependencies and indirect interactions between customers (observations). In this context, graph-based and deep learning models emerge as promising alternatives by enabling the explicit incorporation of customer relationships into the modeling process. However, the construction of these graphs often depends on heuristic criteria, based on the knowledge of business experts or on descriptive analyses performed after model training, such as graphs derived from explicit interactions between customers. Additionally, in many business contexts, there is neither a clear heuristic nor a direct source of customer relationships, such as in companies that sell products digitally, making the definition of relational graphs a significant methodological challenge. This dissertation proposes, implements, and evaluates a comprehensive framework for churn prediction that integrates causal discovery between explanatory variables, construction of customer graphs based on similarity, and predictive modeling via graph-based deep neural networks. The effectiveness of the proposed model is comprehensively evaluated and compared to both variants with random variable selection and state-of-the-art supervised learning models, including Random Forest, Logistic Regression, XGBoost, and LightGBM. The results demonstrate that the proposed approach consistently outperforms baseline models in critical churn metrics, such as true positive rates in low false positive regimes. The findings demonstrate that the framework is appropriate for handling the complexity inherent in the churn prediction problem, showing consistent performance gains and greater structural coherence compared to traditional approaches.
Examination Board
Headlines:
| Marcelo da Silva Reis | IC / UNICAMP |
| Sandra Eliza Fontes de Avila | IC / UNICAMP |
| Ronaldo Cristiano Prati | CMCC / UFABC |
Substitutes:
| Rafael de Oliveira Werneck | IC / UNICAMP |
| Ariane Machado Lima | EACH / USP |