Customer feature selection from high-dimensional bank direct marketing data for uplift modeling

Uplift modeling estimates the incremental impact (i.e., uplift) of a marketing campaign on customer outcomes. These models are essential to banks' direct marketing efforts. However, bank data are often high-dimensional, with hundreds to thousands of customer features; and keeping irrelevant and redundant features in an uplift model can be computationally inefficient and adversely affect model performance. Therefore, banks must narrow their feature selection for uplift modeling. Yet, literature on feature selection has rarely focused on uplift modeling. This paper proposes several two-step feature selection approaches to uplift models, structured to cluster highly relevant, low-redundant feature subsets from high-dimensional banking data. Empirical experiments show that fewer features in a selected set (20 out of 180 features) lead to 68.6% of these uplift models performing as well or better than complete feature set models.