Hempel
Near real-time customer segmentation
Machine Learning & AIManufacturing & Industrials
Key result140,000 customers segmented from 9 data sources
Challenge
Create an intelligent segmentation of the customer base into relevant, coarse segments — with a near real-time continuous learning process — where each segment has a unique profile, meaningful size and stability over time.
Solution
Near-360° data on ~140,000 consumer and business customers, clustered with a robust algorithm and refined iteratively.
- Data sources. Five internal sources (ERP, CRM…) and four external (regional statistics, weather…) — 37 data items in total.
- Algorithm. Partitioning Around Medoids (PAM) clustering — similar to k-NN but less sensitive to outliers.
- Iteration. Model outputs reviewed repeatedly, with adjustments based on intermediate results.
Results
- Private vs. corporate customers reliably differentiated
- Geography-based segments discovered
- Behavioural traits shown to dominate purchasing power & demographics
Topics
- Machine Learning
- Clustering
- AWS