M.S. Applied Data Science - Capstone Chronicles 2025

26

Furthermore, this study extends the work of Bose and Kim (2025). While they successfully applied graph-based segmentation to business-to-business financial transactions, our analysis demonstrated that similar techniques leveraging graph topology can be applied to a business-to-consumer (B2C) network provided that geospatial filtering was utilized. While this study lays the groundwork for future analysis, we acknowledge several inherent limitations: temporal bias, sample representation, and assumptions of brand/location homogeneity. On the point of temporal bias, this analysis relied on a single, cross-section of July 2025 data. Consequently, this means all downstream analysis did not account for seasonal shopping patterns such as holiday shopping or tourism spending. Secondly, as mentioned in the section on data quality, the SafeGraph panel is under-indexed for California by about 1.37% and was likely further biased due to filtering only for San Diego zipcodes. This bias meant that observations were only relevant to the region under observation. In terms of the assumption of brand homogeneity, during construction of the graph’s edges, we made the assumption that given multiple franchises of a single brand, all inbound cross-spending weights would be attributed to the largest absolute percentage among franchises. This assumption was made due to the fact that all cross-shopping data per location was only available at the brand level. Finally, global community detection results were of poor quality based on both Louvain and Laiden community modularity scores. While geospatial filtering did improve clustering, it still did not produce meaningful results, indicating that perhaps our data was not adequately prepared for community detection. Despite these shortcomings, we were able to successfully produce ranked business recommendations based on a custom scoring function that took into account cross-shopping weight, distance to the node of interest, and customer flow similarity.

234