Retail Store Network Optimization with Geospatial ML

Problem

The store network had grown over years without a consistent, data-driven way to match store size and staffing to local demand. Some stores sat too close together and were under-used, while some high-demand areas had no coverage.

Approach

We broke the question into measurable pieces: how far customers actually travel, how close two stores can be before they cannibalize each other, and how much demand each area generates. Those answers fed a demand model, an integer program for the right number of stores per city, and a genetic algorithm that picks an action for every store under cost and coverage constraints.

How It Works

1. 1
   Catchment analysisHaversine distance between 2.4M store visits and customers’ home or work locations gave a typical catchment of about 3.3 km.
2. 2
   Spacing testMann-Whitney U tests on store utilization showed stores within 4 km of a neighbor perform worse, which became the minimum spacing rule.
3. 3
   Grid featuresSpark and Sedona aggregate customers, points of interest, demographics, and anonymized mobility signals onto H3 hexagons; resolution 8 (~0.74 km²) was chosen through entropy analysis.
4. 4
   Segment visitorsK-means on PySpark groups store visitors into archetypes, showing how much demand could move to digital and self-service channels.
5. 5
   Demand modelXGBoost predicts visits per grid cell; an integer program (PuLP) sets how many stores of each size every city needs.
6. 6
   Optimize actionsA multi-objective genetic algorithm (PyGAD) chooses keep, upsize, downsize, close, or relocate for each store; road-distance routing then resolves near-duplicate stores and sizes staffing.
7. 7
   Explain & publishAn LLM writes a plain-language justification for each store in a Kubeflow pipeline tracked with MLflow, and results load into PostGIS for the client dashboard.

What I Built

• K-means visitor segmentation on PySpark.
• LLM-generated per-store justifications, orchestrated on Kubeflow with MLflow.
• Contributions to the XGBoost visit-demand model and the network optimization (integer linear programming with PuLP and a multi-objective genetic algorithm).

Impact

The recommendations went to the client as two scenarios worth roughly 4% and 9% in annual cost efficiency while keeping catchment coverage. Decision thresholds were checked against the business’s own past decisions.