The business problem
Call center capacity is expensive. When approaching customers with a new add-on product, our client so far has been using a random selection model, but this has yielded very low (10-15%) efficacy. Can this be optimized better so as to reach the most susceptible customers?
A complex model based on recursive partitioning used data about customers and their behavioral patterns to predict which customers are most likely to respond positively to an approach. Additional locality-based and interaction patterns were extracted from customer data.
92% model accuracy (as measured by ROC AUC) and a very significant 6x uplift in conversion rate