Feature Selection Methods Compared for Customer Targeting.

Feature selection is a critical step in data-driven projects, aiming to identify the most influential variables for predicting an outcome. This research conducts a comparative study of two prominent feature selection techniques: mutual information and data-based sensitivity analysis. The comparison is performed in the context of a bank telemarketing campaign, where the goal is to predict the success of a customer contact. After applying both methods, a logistic regression model was built using the features identified by each. Mutual information selected 13 features, while data-based sensitivity analysis identified 9. The results showed that data-based sensitivity analysis performed better when minimizing false positives was the priority. In contrast, mutual information proved slightly superior for scenarios where a higher false positive ratio was acceptable. The study concludes that mutual information, despite being an older technique, remains a valid and effective method for feature selection, particularly if the objective is to slightly reduce contact costs without risking a significant loss of successful outcomes. Data-based sensitivity analysis, while using fewer features, also achieved good prediction results, offering a more parsimonious model.