Client segmentation under real-world constraints

Abstract

Market segmentation is a very useful tool that can enhance knowledge of a firm's customer base and therefore enable improved customer services and experiences that are more tailored to specific customer needs and preferences. Clustering is a natural and intuitive way to implement such segmentation, and in fact, there are a variety of standard methods by which to perform this. However, real-world considerations complicate its implementation, in particular, the necessity of not clustering in ways that could be considered discriminatory in terms of certain features such as gender or race. One way to mitigate such discriminatory clustering is through constraints that ensure that the clusters are balanced in terms of such features. However, such a clustering is barely, if at all, discussed in current literature. In this thesis, we develop and implement a new version of k-means clustering that is able to achieve comparable performance relative to an unconstrained clustering while at the same time address the constraints imposed by these discriminatory features.