Insight Project

Insight project: High value customer profile extration

Transcript

1. High Value Customer Profile Extrac6on Miao Xie Consul6ng Project
2. : Help business owners discover loca6ons Frozen Food Discovered location
3. : Where do valuable customers come from? Extract valuable customer profiles
4. Data • 3.5 million customers • Transac6on, geoloca6on Customer • 300+ stores • Annual sales, geoloca6on Store • 50K neighborhoods • Demographics, geoloca6on Neighborhoods
5. Data • 3.5 million customers • Transac6on, geoloca6on Customer • 300+ stores • Annual sales, geoloca6on Store • 50K neighborhoods • Demographics, geoloca6on Neighborhoods
6. Data • 3.5 million customers • Transac6on, geoloca6on Customer • 300+ stores • Annual sales, geoloca6on Store • 50 K neighborhoods • Demographics, geoloca6on Neighborhoods Valuable customers
7. What metric iden6fies valuable customers?
8. What metric iden6fies valuable customers? No metric exists!
9. Sales What metric iden6fies valuable customers? No metric exists!
10. Sales What metric iden6fies valuable customers? No metric exists!
11. The Metric: Customer AKrac6on Score
12. The Metric: Customer AKrac6on Score How a&racted customers are to a store Customer Attraction Score = Store Sales Nearby Population
13. Customer Attraction Score = Store Sales Nearby Population Frozen Food
14. Frozen Food Customer Attraction Score = Store Sales Nearby Population Neighborhoods
15. Popula6on Size 4 4 1 3 5 1 18 Neighborhoods Customer Attraction Score = Store Sales Nearby Population
16. Data loaded in PostgreSQL Data cleaning AKrac6on score (AS) Super store Average store Features Model Logis6c regression Random forest classifica6on Work Flow Scikit-learn KMeans cluster Missing values, outliers
17. Data loaded in PostgreSQL Data cleaning AKrac6on score (AS) Super store Average store Features Model Logis6c regression Random forest classifica6on Work Flow Scikit-learn KMeans cluster Missing values, outliers
18. Data loaded in PostgreSQL Data cleaning Customer aKrac6on score Super store Average store Features Model Logis6c regression Random forest classifica6on Work Flow Scikit-learn KMeans cluster
19. Data loaded in PostgreSQL Data cleaning Customer aKrac6on score Super stores Average stores Features Model Logis6c regression Random forest classifica6on Work Flow KMeans cluster
20. Data loaded in PostgreSQL Data cleaning Customer aKrac6on score Super stores Average stores Features Model Logis6c regression Random forest classifica6on Work Flow
21. Who are the valuable customers?
22. People without a degree like the store 0 0.1 0.2 0.3 0.4 # No degree $ Vegetable Age > 65 yr Feature importance from random forest model
23. People with large spending in vegetables are less likely to visit the store 0 0.1 0.2 0.3 0.4 # No degree $ Vegetable Age > 65 yr Feature importance from random forest model
24. Seniors like the store 0 0.1 0.2 0.3 0.4 # No degree $ Vegetable Age > 65 yr Feature importance from random forest model
25. Validation Model to predict high value store loca6ons Random forest classifica6on
26. Complete overlap Predicted VS Current Store Locations
27. New locations Predicted VS Current Store Locations
28. Valida6on: AUC = 0.92 and Accuracy = 87%
29. Miao Xie PhD in Physical Chemistry at UCLA –  Characterizing superhard materials Product engineer in Silicon Valley
30. Backup