Clustering of retail network object assortments based on Jaccard distance and dendrograms

Abstract

This paper addresses the issues of automated clustering of product assortments in retail networks. The study is particularly relevant for retail chains with a large number of stores and a broad assortment exceeding 2,000–3,000 items. Under such conditions, comparative analysis of effective assortments requires substantial time and financial resources, as it is typically performed manually by analysts or with a limited level of automation. Optimization of assortment management processes can be achieved by grouping stores into homogeneous clusters, which simplifies managerial decision-making, enables standardization of assortment policies, and improves branding efficiency. Traditionally, store classification is based on sales area size and total assortment volume. However, these criteria do not always reflect the actual similarity of product offerings. This paper proposes an algorithm for automated store clustering based on the analysis of assortment overlap. The method is based on constructing a binary matrix representing product availability across stores and computing pairwise Jaccard distances between store assortments. Using the resulting distance matrix, hierarchical clustering is performed with dendrogram construction, allowing the determination of an optimal number of clusters according to predefined parameters. Additionally, the clustering results are visualized using the t-distributed Stochastic Neighbor Embedding (t-SNE) method, which provides an intuitive interpretation of the data structure in a reduced-dimensional space. The proposed approach enables not only the formation of store groups with a common core assortment but also the identification of product items that distinguish individual stores within the retail network. The obtained results can be applied to optimize assortment matrices, design effective store formats, and reduce assortment management costs in large retail chains.