2. 3-12-2014 pag. 2
Who?
• Professor of Business Informatics & Business Analytics
– Faculty of Economic and Social Sciences and Solvay Business School, VU Brussels, Belgium
– Associated with BUTO, MOBI, SMIT - iMinds
– Business Analytics Team: 4 Ph.D. students
• Research topics:
– Business Analytics: focus on business user requirements
• End use optimization: ROI, profit, … optimizing analytics
• Comprehensibility
• Justifiability
• …
– Application driven research
• Credit risk management
• Customer relationship management
• Demand forecasting
• Fraud detection
• …
– Network analytics
5. 3-12-2014 pag. 5
Marketing analytics
Operations
Data
Decision
making
Evaluation
6. 3-12-2014 pag. 6
• Correct evaluation
– Evaluation measure
Verbeke W., Dejaeger K., Martens D., Hur J., Baesens B., New Insights into Churn Prediction in the Telco Sector: a Profit
Driven Data Mining Approach, European Journal of Operations Research, 218 (1), pp. 211–229, 2012
7. Optimizing return of retention campaigns
Churners
3-12-2014 pag. 7
• Customer churn and retention:
dynamics within customer base
New
customers
• Return of a retention campaign:
Π=푁훼{훽훾(퐶퐿푉−푐−훿)+ 훽(1−훾)(−푐)+(1−훽)(−푐−훿)}−퐴
Outflow
Customer
base
(size N)
True would-be
churners
(β)
False would-be
churners
(1-β)
Inflow
Nα customers included in a
retention campaign and
offered an incentive (δ)
All retained
Fraction γ
retained
Fraction 1-γ not retained
8. • The term β represents the ability of the model to identify would-be
3-12-2014 pag. 8
churners, and 훽 = 훽0휆(훼):
Π = 푁훼 [훾퐶퐿푉 + 훿(1 − 훾)]훽0휆(훼) − 훿 − 푐 − 퐴
• The maximum profit measure is defined as:
MP = max
훼
(П)
• Managerial implications: 훼표푝푡푖푚푎푙
• Benchmarking study: significant profit gains!
9. 3-12-2014 pag. 9
• Correct evaluation
– Evaluation measure
– Evaluate campaign effect: control groups
10. 3-12-2014 pag. 10
• Control groups: campaign measurement of model
effectiveness
Treatment Control
Target
group
Non-target
group
Random
target
group
Random
non-target
group
Model
Random
Lo, V., The true lift model – a novel data mining approach to response modeling in database marketing
11. 3-12-2014 pag. 11
Operations
Data
Decision
making
Evaluation
Evaluation
data
12. 3-12-2014 pag. 12
• Netlift modeling:
age usage trend gender targeted churn
21 +5,2 M Yes Yes
37 +0,1 F Yes No
19 -3,2 U Yes No
45 +4,2 F No Yes
28 +2,1 M No No
62 -2,3 F No No
… … … … …
17. 3-12-2014 pag. 17
Go further
• Objective function = evaluation criterion?
– Cost-sensitive learning
• At the class level: cost of misidentification
• At the individual level
– Customer lifetime value
– Balancing
18. 3-12-2014 pag. 18
Go further
Christakis, Nicholas; Fowler, James H. - Dynamic Spread of Happiness in a Large Social Network: Longitudinal Analysis Over 20 Years in the
Framingham Heart Study (http://dash.harvard.edu/handle/1/3685822)
19. 3-12-2014 pag. 19
Social network analytics
• Social network analysis for customer churn prediction
– Featurization, propositionalization, …
– Network analytics:
• Predictive
– Relational learning
– Viral approaches
– Graph based
• Descriptive
– Centrality measures
– Link, node, degree distributions
Verbeke W., Martens D., Baesens B., Social network analysis for customer churn prediction, Applied Soft Computing, 14, pp. 431-446, 2014
21. 8 7 2 3
9 7 2
4 3 2
3 2 9
3-12-2014 pag. 21
Relational learning
181806208300809 32462208699 206105300897975 357014032645640 I 32461002530 9 MOBISTAR MOBILE 99 21JAN2010:23:45:44 0 0 0 0 2 1 1 …
195455641 32475611232 206102200262341 351913035725230 I 32476000005 10 Base SMSC Platform 99 21JAN2010:23:46:02 0 0 0 0 2 1 1 …
187097451101277 32465245451 206101100499483 356712034636630 I 32473161616 8 Proximus SMSC Platform 99 21JAN2010:23:45:44 0 0 0 0 2 1 1 …
Raw
CDRs
21
8 9 4 3
21
3
2
3 3
2 3 8
9 8
A
B
C
D
E
F
G
H
I
J
A B C D E F G H I J
Sparse
connectivity
matrix
C
A D
E
F B
J
H
I
G
Weighted
network
8
9
4
3
2
3
3
3
2 2
9
8
7
22. 3-12-2014 pag. 22
Findings
• Predictive power?
• Featurization?
• Combine with traditional model?
23. 3-12-2014 pag. 23
Go further
• Survival analysis
• Survival analysis …
… with netlift modeling
Thomas
Marc
Loyalty
Time
Marketing
campaign #1
Loyalty
Time
Thomas
Marc
Marketing
campaign #2
24. 3-12-2014 pag. 24
What did we learn today?
• Checklist
– Evaluation
– Let’s test
– Bring in the next steps
– It’s the data, stupid!
– Parallel models
– Towards real-time customer tracking