Project Writeup

CS Department Presentation & Marketing Department Presentation

I employed a unique configuration of the partial clustering algorithm, Mapper, to visually analyze and identify classification errors between various statistical or machine learning prediction methods. Using proprietary customer data, I applied various customer dropout models including deep neural nets, support vector machines, and gradient boosted decision trees along with traditional statistical prediction models (e.g., Fader et al., 2005). I then determined a unique use of partial clustering for error detection, visually compared the prediction error of individual customers across models partially clustered by their consumer features and demographics (model inputs), and discovered a unique segment of edge case customers that are not responsive to traditional models but have mixed results with learning-based models. 

Available Upon Request

As machine learning approaches (such as neural nets) become commonplace in customer estimation, deeply embedded information such as gender is more likely to be teased out and implicitly or explicitly included in the model calculation. We examine six statistical and machine learning customer dropout models that either implicitly or explicitly consider gender using false negative and false positive parity. In a data collaboration with a retail firm, we find that men are both more likely to be misclassified as alive and less likely to be misclassified as dead compared to women. Both phenomena directionally make men more likely to be added to customer retention programs regardless of dropout model, or explicit inclusion of gender into that model.