Persistence of Preference- Based Customer Segments : An investigation of cluster evolution

Abstract: Clustering is a technology within unsupervised learning with a wide range of applications. Several of these applications use data that change over time, which makes clusters’ persistence of interest. One among these employments of clustering time-variant data is preference based customer segmentation. Preferences are assumed to change over time and it is thus of interest to know for how long clusters based on preferences remain. This study explores clusters of clients obtained in the segmentation analysis of users of a video streaming service and their persistence over time. The clients were clustered based on viewing history from distinct months with the k-means algorithm. Various metrics, such as Rand Index (RI), Adjusted Rand Index (ARI) and Fowlkes-Mallows score, were employed for evaluation of cluster persistence. It was found that most of the identified clusters did not show persistence over months but that most partitions included at least one clustered that was considered persistent. The results also suggested that clusters featured by titles that target children were more persistent than other clusters. Moreover, clients with a large interest in videos within the children genres appeared to form relatively separated clusters, which supports considering consumers of children titles as a separate target group.