Mapping Regional Disparities in Discounted Grocery Products

Abstract

Food waste represents a major challenge to global climate resilience, accounting for almost 10\% of annual greenhouse gas emissions. The retail sector is a critical player, mediating product flows between producers and consumers, where supply chain inefficiencies can shape which items are put on sale. Yet how these dynamics vary across geographic contexts remains largely unexplored. Here, we analyze data from Denmark's largest retail group on near-expiry products put on sale. We uncover the geospatial variations using a dual-clustering approach. We identify multi-scale spatial relationships in retail organization by correlating store clustering -- measured using shortest-path distances along the street network -- with product clustering based on promotion co-occurrence patterns. Using a bipartite network approach, we identify three regional store clusters, and use percolation thresholds to corroborate the scale of their spatial separation. We find that stores in rural communities put meat and dairy products on sale up to 2.2 times more frequently than metropolitan areas. In contrast, we find that metropolitan and capital regions lean toward convenience products, which have more balanced nutritional profiles but less favorable environmental impacts. By linking geographic context to retail inventory, we provide evidence that reducing food waste requires interventions tailored to local retail dynamics, highlighting the importance of region-specific sustainability strategies.

Figures (9)

• Figure 1: Spatial patterns of retail product on sale distribution across Denmark's administrative regions.a) Map of Denmark showing the five administrative regions analyzed: Hovedstaden (red), Sjælland (pink), Syddanmark (teal), Midtjylland (light blue), and Nordjylland (blue), with store locations and spatial clustering results overlaid. Map data copyrighted OpenStreetMap contributors and available from https://www.openstreetmap.org. b) Hierarchical spatial clustering analysis showing the core-periphery structure of retail locations based on shortest-path distances along the street network. The clustering groups part of Sjælland and all of Hovedstaden into one cluster, identifies a mix of Midtjylland and part of Syddanmark, captures another part of Syddanmark in a separate cluster, and forms clusters corresponding mainly to Nordjylland and Sjælland, while noise is scattered primarily across Jutland. c) Regional analysis of product-on sale category frequencies and diversity patterns. The lollipop chart, corresponding to a rank plot, shows commonly on sale product categories nationwide (Cheese, Pork, Ready-to-Eat Meals, Breads, and Yoghurt), while the inset of normalized entropy highlights regional differences in product diversity. d) Mantel correlation analysis (green) examines the relationship between spatial proximity and promotional composition similarity across multiple distance scales. The correlogram reveals three distinct phases (grey vertical lines): (I) positive spatial-product correlations at smaller scales peaking around $4000$ km, driven by the cohesive Hovedstaden cluster; (II) correlation diminishment upon inclusion of the Sjælland region, with the inversion of the Mantel correlation (and becomes statistically non-significant as highlighted by the shaded area) occurring at the point where the second-largest cluster (orange) merges with the first, as tracked by the blue line (fraction of nodes in largest cluster divided by the maximum value); and (III) negative correlations at larger scales ($\sim26500$ km) indicating compositional divergence among distant stores.
• Figure 2: Bipartite network analysis identifies product-driven regional retail differentiation in Denmark.a) The infographic summarizes the bipartite network with stores and products on separate layers. b) Store-layer projection after validation showing three well-defined communities corresponding to Metropolitan Areas (red), Capital Region (blue), and Countryside (green) clusters. The nodes are positioned according to their geographic coordinates (latitude and longitude), while avoiding overlaps. Node size proportional to the logarithm of the population. c) Fraction of nodes in the largest cluster (red) and the second-largest cluster (teal) identified via percolation analysis, with clusters merging at the same distance indicated by the hierarchical clustering (grey vertical line $\sim 26500$ km). d) Products-layer projection using the Yifan Hu layout. Node colors highlight different categories and size proportional to the number of neighbors. The network contains 124 communities, most of which are actually disconnected components (111). e) Rank‐ordered probability of each product category appearing among the top 10% of nodes by betweenness centrality.
• Figure 3: Community-specific promotional composition profiles and prevalence of unhealthy products in peripheral danish communities.a) Per-capita on sale frequency (per 100,000 potential customers) for product categories (color) across Metropolitan Areas, Capital Region, and Countryside. The Cheese category has been removed, as its values exceed 0.6 across all communities, and values below 0.01 have also been excluded to improve readability. b) Distribution of Nutri-Score ratings (A: healthiest to E: least healthy) across the three community types, showing prevalence of categories B and D and under-representation of the healthiest products in all regions. c) Environmental Score distribution (A+: lowest impact to E: highest impact) reveals mid-grade products (B) dominating across communities, with Countryside showing more environmentally positive options while Metropolitan Areas and Capital Region lean toward less environmentally favorable products.
• Figure S1: Dendrograms of stores based on Bray–Curtis similarity. Each leaf represents a store, with adjacent colored bars denoting the corresponding region: Hovedstaden (red), Sjælland (pink), Syddanmark (teal), Midtjylland (light blue), and Nordjylland (blue). We observe a tendency for stores belonging to the same region to cluster more closely within the dendrogram.
• Figure S2: Robustness analysis of Mantel correlogram under incomplete data scenarios. Mantel correlogram based on incomplete datasets (with 5, 10, and 15 randomly removed days). The standard deviation is not shown, as it is negligible ($<10^{-4}$).