User segmentation via interpretable user representation and relative similarity-based segmentation method

Abstract

User segmentation is an essential element of marketing and product development that considers customers’ needs and recognizes the heterogeneity of those needs. In a key study of smartphone user segmentation, Lee et al. analyzed app usage sequencing using seq2seq architecture. However, despite achieving meaningful results, their approach could not provide a robust interpretation of user segmentation because the seq2seq architecture represented users in a continuous vector space generated from a black box model. In this paper, we propose an interpretable user representation method that combines app clustering with a novel segmentation method. The user representation clusters characteristically similar apps into common clusters, with each user represented by their frequencies of app use within their respective clusters. Two novel techniques are also applied to normalize the value of user representation based on the relative degrees of importance between app clusters and the membership strengths of individual apps within a cluster. Furthermore, to address the limitations of existing segmentation methods, in which the most closely located users are assigned to specific clusters, the proposed method segments represented users using a novel segmentation approach based on relative similarity. Experimental results demonstrate that the proposed method provides an intuitive interpretation for each user’s representation and segmentation results. Furthermore, we effectively show the similarities between the results produced by our method and ground truth and demonstrate that it outperforms existing user segmentation methods.