BADM Project Hyper Market Classifying Biscuit Brand Switchers for - - PowerPoint PPT Presentation

Group B4 - Minesweepers Aditi Vaish | Pranav Maranganty | Kevin John | Deepak Agnihotri | Archana Rajan BADM Project – Hyper Market Classifying Biscuit Brand Switchers for Targeted Marketing for a Biscuit Manufacturer

Business Problem

• Client Profile and Background
• Minesweepers Biscuits (MSB) based out of Denmark
• Renowned brand internationally but limited brand presence in India
• Expansion to India with new products and innovative promotions
• Competition
• Brands like Britannia and Parle own a big pool of loyalists
• Business Objective
• Increase Trial Rate of MSB products and hence Market Share by

partnering with Hyper Market

• Improve Marketing Efficiency by targeting only the “Brand Switchers”
• Offer samples and personalized promotions at checkout counters or at

kiosks of the Hyper Market to the “Brand Switchers”

All Consumers Selected Consumers for better ROI on marketing

2000 4000 6000 8000 10000

Number of Enrollments

Data Mining Problem

• Objective
• Classify new customers as “Brand Loyalist” (1) and “Brand Switchers”

(0) based on demographics and purchase patterns in “Ready Food”

• New Customer – A customer who has made 2 purchases from the

“Ready Food” department

Business Rules Classification

Continuous Evaluation and Classification

MSB’s Target “Brand Loyalist”: A person who has purchased biscuits at least 3 times and purchased the same brand over 50% of the times

Data

• Key Inputs ( Vary depending on the various models)
• Output
• Loyalists? (0 – Brand Switcher , 1 – Brand Loyalist)

Customer Demographics Historical Purchase Pattern ( Ready Food) Last Purchase (Ready Food) Second Last Purchase (Ready Food) Age Sex Marital Status Enrollment Store Average Basket Price Average Basket Quantity Average Basket Unique Count Number of Baskets Standard Deviation of Basket Price Standard Deviation of Basket Quantity Quantity Price Unique SKU Count Quantity Price Unique SKU Count Data Partition Training Set : 2172 Validation Set : 1303 Test Set: 868 Hold Out for Model Evaluation: 500 Data Total Unique Customer: 4843 Brand Loyalists : 2886 (59.6%) Brand Switchers: 1957 (40.4%) A person who has purchased biscuits at least 3 times Initial Classification: A person who has purchased biscuits at least 3 times and purchased the same brand over 50% of the times

Methods

CART Pruned Tree : 3 Nodes # Input Variables: 18 Naïve Bayes # Input Variables : 6

Class # Cases # Errors % Error 524 208 39.69 1 779 334 42.88 Overall 1303 542 41.60 Class # Cases # Errors % Error 905 241 26.63 1 1267 456 35.99 Overall 2172 697 32.09

Training Set Test Set

Class # Cases # Errors % Error 359 146 40.67 1 509 235 46.17 Overall 868 381 43.89

Validation Set

K-NN Best K: 5 # Input Variables : 14

Class # Cases # Errors % Error 905 112 12.38 1 1267 613 48.38 Overall 2172 725 33.38 Class # Cases # Errors % Error 524 148 28.24 1 779 508 65.21 Overall 1303 656 50.35 Class # Cases # Errors % Error 359 95 26.46 1 509 313 61.49 Overall 868 408 47.00

Value of k % Error Training % Error Validation 1 0.00 44.67 2 23.94 49.12 3 23.16 45.13 4 27.53 47.51 5 28.68 43.13 6 30.66 47.12 7 30.34 43.90 8 31.72 45.89 9 32.27 43.75 10 32.97 44.74

Probability Cutoff: 0.4

Class # Cases # Errors % Error 905 0.00 1 1267 0.00 Overall 2172 0.00 Class # Cases # Errors % Error 524 189 36.07 1 779 321 41.21 Overall 1303 510 39.14 Class # Cases # Errors % Error 359 125 34.82 1 509 215 42.24 Overall 868 340 39.17

Logistic Regression - Stepwise Initial No. of Variables: 21 # Variables based on Cp: 20

Class # Cases # Errors % Error 905 386 42.65 1 1267 444 35.04 Overall 2172 830 38.21 Class # Cases # Errors % Error 524 222 42.37 1 779 272 34.92 Overall 1303 494 37.91 Class # Cases # Errors % Error 359 142 39.55 1 509 198 38.90 Overall 868 340 39.17

Coefficient

• 47.1890297

0.00544622 1.11986947 1.03916395

• 0.67768991

0.1495695 0.09804565 0.07624547

• 0.05565267
• 0.00135054
• 0.04568335

0.00852676 0.00019505 0.02833186 0.00110265 0.01369065 0.00012139

• 0.00228719
• 0.00255478

0.00014995 Input variables Average Basket Quantity Constant term Age Sex_F Sex_M Enrollment Store_1001 Enrollment Store_1002 Marital Status_N Marital Status_Y Email_Y Average Basket Price Number of Baskets StdDev of Basket Price StdDev of Basket Quantity Last Transaction Date Last Purchase Unique Count Last Purchase Price Last Purchase Quantity Second Last Purchase Second Last Purchase Price

We tried different # variables, but all had some over fitting

Model Evaluation (Test Data)

• Naïve Rule is considered as the benchmark

for evaluation with all customers tagged as “Brand Loyalist”

• Key Metrics for Evaluation
• Sensitivity (0  1)
• All models are better than the benchmark
• % Total Error
• Logistic Regression and CART fare better than

benchmark

• Misclassification Costs
• INR 120 for 0  1 (Customer Value)
• INR 20 for 1  0 (Coupons and Samples)
• All models fare better than benchmark
• Holdout Evaluation - K-NN has some over fitting

100.00% 0.00% 41.36% 0.00% 20.00% 40.00% 60.00% 80.00% 100.00%

Sensitivity Specificity % Total Error

Benchmark Logistic Regression Naïve Bayes K-NN CART

61080 21000 22220 17660 19300 20000 40000 60000 80000

Benchmark Logistic Regression Naïve Bayes K-NN CART

Logistic Regression

Class # Cases # Errors % Error 190 86 45.26% 1 310 102 32.90% Overall 500 188 37.60% Class # Cases # Errors % Error 190 144 75.79% 1 310 73 23.55% Overall 500 217 43.40%

CART

• Deploy Logistic Regression Model for classifying new customers
• Low Misclassification Costs
• Similar Accuracy across all Data
• Better Overall Error and Sensitivity
• Easy Deployment of Model / Stable Model
• Continuously improve the model by updating the classifications

and adding more data

• Include External Demographics data to improve the model
• Expand model to include other products of MSB

Recommendations

Next Steps