Essays in retailing and distribution

• Main Author: Li, Tieshan
• Language: English
• Published: University of British Columbia 2009
• Online Access: http://hdl.handle.net/2429/7505

Although aggregating retail outlets into retail districts is an important academic and practical issue in marketing and retailing, only limited academic work has been done on this problem. The growing availability of detailed location data through Geographic Information Systems makes this a particularly timely problem. Cluster analysis is a sound and well established approach for reducing data dimensionality. However, the existing clustering approaches do not handle the complicated geospatial structure that is typical of retailing data well, largely due to the high variation in observation density. One problem is that the “epsilon radius,” a measure of how close stores need to be to each other in order to be classified as belonging to the same cluster, is assumed to be constant in methods such as density-based clustering. However, this turns out not to be a good assumption in practice. In addition existing methods of judging the quality of a clustering solution, so called cluster validation methods, do not provide sound guidance as to the best clustering solution for the type of retailing data we study. Consequently, we propose a new two-step clustering approach in which Variable Epsilon Spatial Density Clustering (VESDC) is developed, and a new clustering validation measure, the CpSp index, also is introduced. VESDC effectively clusters data by systematically adjusting the epsilon radius to adapt to the local market environment. In particular, using the logistic transformation function, we propose a model in which the epsilon radius is determined by the population density in a small area. CpSp, which is scaled from 0 to 1, balances the compactness and separation of a proposed clustering solution. Extensive testing demonstrated that CpSp performed well as a cluster validation method. We tested VESDC’s performance on synthetic data. The underlying pre-specified data patterns were accurately recovered. Existing methods were not as successful in these tests. We then applied the two-step approach to Greater Victoria since Greater Victoria is a typical metropolitan city with large variation in store density.