FeatureSeries: An explainable neural network for financial marketing

Financial marketing models often struggle with interpretability due to the complexity and high dimensionality of features. This paper introduces FeatureSeries, an explainable neural network designed to address these challenges in financial marketing scenarios. Our model tackles two primary issues: feature selection from vast repositories and effective feature ranking. To improve feature selection, we incorporate an orthogonality loss function during model training, which mitigates collinearity among selected features. For feature ranking, we innovatively apply Wasserstein distance combined with Sobol coefficients, enabling more accurate importance assessment of behavioral and potentially redundant features common in financial marketing. FeatureSeries demonstrates superior performance in both predictive accuracy and model explainability compared to existing methods. We optimize Generalized Additive Models (GAMs) through our novel approach, resulting in more interpretable output features and enhanced model performance. Our experiments on real-world financial marketing tasks, including customer segmentation and product recommendation, validate the effectiveness of FeatureSeries. This paper contributes to the field by presenting a model that not only improves the interpretability of complex financial marketing models but also maintains high predictive power. The insights provided on applying FeatureSeries to various financial marketing scenarios offer practical value for industry professionals and researchers alike.