Stay Classy: Maximizing Promotion Returns Using Classification · Stay Classy: Maximizing Promotion Returns Using Classification Katherine Sanborn Manager, Business Analytics Kellogg

#AnalyticsXC o p y r ig ht © 201 6, SAS In st i tute In c. A l l r ig hts r ese rve d.

Stay Classy: Maximizing Promotion Returns Using ClassificationKatherine SanbornManager, Business AnalyticsKellogg Company

Project beginnings

• Exploratory project utilizing EM to understand underlying patterns in our sales data

• Objective is to determine promotion attributes important to incremental lift

• Typically would use time-series forecasting with this data, but are exploring new data and tool uses

2

Agenda

• Data

• Methodology

• Visual Analytics Exploration

• Modeling Results

• Accuracy Assessment

• Lessons Learned

• Conclusion

3

Data

• Three years of point of sale data

• Our products and major competitive products

• Incremental and base variables including pricing, merchandising, and volumetric measures

– Merchandising includes display and feature measures

4

Methodology

• Define incremental lift

– Base sales

– Incremental sales

• Start with three groups segmented by incremental lift

• Use SEMMA as process to build and assess models

– EG, EM, and Visual Analytics used in various stages of project

5

Visual Analytics Exploration

6

Models from RPM Diagram

7

Custom EM Models from Project

8

Accuracy Assessment

Which is better?

(A.) Model 1: 85% Accuracy (15% misclassification)(B.) Model 2: 70% Accuracy (30% misclassification)(C.) Model 3: 75% Accuracy (25% misclassification)

Actual Classification

Model 1's Classification

Model 2's Classification

Model 3's Classification

Class X: 80 items 95 items 50 items 55 items

Class Y: 20 items 5 Items 50 items 45 items

Model 1: Assign almost

everything to X

Model 2: 50/50 Chance Assignment

Model 3: Decision Tree Assignment

X Y X Y X Y

X 80 0 X 50 30 X 55 25

Y 15 5 Y 0 20 Y 0 20

Model 3

Act

ual

75%

Act

ual

Model 1

85%

Model 2

Act

ual

70%

9

Accuracy Assessment

Cohen’s Kappa & Cohen’s Weighted Kappa

• Takes into consideration chance agreement • Allows for weighting more significant misclassifications

KappaAccurac

y

Model 1 1 - 15 / 23 = 0.348 85%

Model 2 1 - 30 /50 = 0.400 70%

Model 3 1 - 25 / 47 = 0.468 75%

It depends on the cost of the misclassification

X Y X Y

95 * 80 / 100 5 * 80 / 100

76.00 4.00

95 * 20 / 100 5 * 20 / 100

19.00 1.00

95 5 100

X Y

X

Y

X

Y

Act

ua

l

Weight Matrix

1

1 0

0

80 0 80

20Y

X

Model 1 Expectation Matrix

15 5

Which model is “better”?

10

Accuracy Assessment

Rapid Predictive Modeler Models: Accuracy Kappa W.Kappa

Main Effects Regression 64.7% 40.8% 47.7%

Forward Selection Regression 64.5% 40.4% 47.2%

Stepwise Regression 64.5% 40.6% 47.4%

Decision Tree 61.7% 39.2% 47.4%

Backwards Selection Regression 65.0% 41.3% 47.9%

Neural Network 65.2% 41.4% 48.3%

Ensemble Champion 65.0% 42.7% 49.0%

Enterprise Miner Workflow Models:

Gradient Boosting 66.0% 42.7% 49.0%

Logistic Regression (no interactions) 68.1% 47.1% 53.3%

Neural Network 68.4% 47.7% 54.2%

Regular Linear Regression 68.2% 47.3% 53.4%

Ensemble Models 68.2% 47.3% 53.6%

Weighted Matrix

A B C

A 0 1 2

B 1 0 1

C 2 1 0

Penalizes misclassification of A C and C A more than others

11

Lessons Learned

• Certain impactful promotional measures drive incremental lift

– Including feature and display variables, and price discounts

• Additional promotion attributes are needed to fine tune classification

12

Conclusion

• You can start data mining with RPM and EM today

• Gathering your data and start with RPM

13

C o p y r ig ht © 201 6, SAS In st i tute In c. A l l r ig hts r ese rve d.

#AnalyticsX