Skip to main content

Mathias Carlsson: Discovering characteristics of customers with distinct profitability using unsupervised learning

Time: Thu 2022-06-09 10.00

Location: Kräftriket, house 5, room 15

Respondent: Mathias Carlsson

Supervisor: Chun-Biu Li

Export to calendar

Abstract:

In this thesis we are interested in testing how modern methods of unsupervised learning can be applied to real-world problems. To do so, a case study from the insurance industry is considered, which is done in collaboration with Dina Försäkring AB. The data set obtained consists of 36 352 customers, all of which have three profitability variables and 49 characterizing variables. As the names suggest, the profitability variables describe how profitable a customer is for the company, while the characterizing variables describe the characteristics of the customer. The goal of the case study is to segment the customers into different profitability groups using the profitability variables, and then, for each group, discover its characteristics from the set of characterizing variables. To achieve this goal, profitability groups are first created by performing an exploratory analysis. Then, the characteristics of the profitability groups are searched for by performing non-linear dimensionality reduction on the characterizing variables and analyzing the resulting embeddings. The results of the analysis show that no distinguishing characteristics can be found for any of the profitability groups. However, in one of the embeddings, a relationship can clearly be seen between a customer’s profitability group and the customer’s location in the embedding. The relationship is given by an increased distance from the origin when decreasing the profitability. This finding suggests that there may in fact be combinations of characterizing variables that distinguish the profitability groups, which can be explored in future studies.