Name: WADHAM ENTRINGER BOTTACIN
Publication date: 05/08/2025
Examining board:
Name |
Role |
|---|---|
| DIEGO ROBERTO COLOMBO DIAS | Examinador Interno |
| GIOVANNI VENTORIM COMARELA | Presidente |
| GLAUBER DIAS GONÇALVES | Examinador Externo |
Summary: The increasing competition among fixed Internet providers in Brazil has intensified the
need for effective customer retention strategies. This study investigates customer churn
based on the analysis of complaints submitted to Anatel by users of a major telecom-
munications company. The main objective is to combine prediction, interpretability, and
natural language generation to support decision-making in automated customer service.
Three predictive models were developed: one using categorical data, another using textual
data, and a hybrid model that integrates both. Interpretability was achieved through Inte-
grated Gradients to identify relevant tokens associated with churn. The results show that
text-based models outperform the others, reaching an F1-score of up to 0.73 for the non-
retained class. Words most associated with churn include “cancelamento” (cancellation),
“reclamação” (complaint), and “serviço” (service), while “despesa” (expense), “perfor-
mance”, and “contenção” (containment) stood out among retained customers. Finally, the
predictive and interpretable results were used as input for small locally-executed SLMs,
which generated personalized responses to the complaints, ensuring the privacy of sensitive
data. These responses were automatically evaluated using GPT-4o, based on criteria such
as empathy, clarity, and technical adequacy. The findings highlight the potential of an
integrated, secure, and interpretable approach to mitigate churn in telecommunications
services.
