Direct Marketing When There Are Voluntary Buyers

1
Direct Marketing When There Are Voluntary Buyers
Presenter _____________

• Yi-Ting Lai, Ke Wang
• Simon Fraser University
• llai2, wangk_at_cs.sfu.ca

Daymond Ling, Hua Shi, Jason Zhang Canadian
Imperial Bank of Commerce Daymond.Ling, Hua.Shi,
Jason.Zhang_at_cibc.com
2
Introduction Direct Marketing

• Target a selected group of customers.
• Which customers should be selected for contact so
  that the campaign can achieve the maximum net
  profit?
• Traditional objective identify the customers who
  are most likely to respond.
• A real direct marketing campaign

Assumption All profits are generated by the
campaign!
5.4
80 are voluntary buyers!
4.3
3
Three Classes of Customers

• Based on their purchasing behaviors
• Each customer belongs to exactly one class

Decided these customers voluntarily buy the product, regardless of a direct promotion.
Undecided these customers will buy the product if and only if the product is directly promoted to them.
Non these customers will not buy the product, regardless of a direct promotion.
The only customers who can be positively
influenced.
4
Is the traditional paradigm solving the right
problem?

• Given a fixed number of contacts, need to
  maximize the set of total buyers in order to
  maximize net profits.

undecided
undecided
M2
The traditional paradigm favors M1.
M1
non
non
decided
decided
undecided
undecided
M2 has targeted more buyers!
M2
non
M1
non
decided
decided
5
Influential Marketing

• S the set of contacted customers.
• DBR the decided buyer rate of S.
• UBR the undecided buyer rate of S.
• RR the response rate of S.
• Influential Marketing
• For a given number of contacts, influential
  marketing aims to maximize UBR by targeting
  undecided customers.
• Challenges
• Customers are not explicitly labeled by the three
  classes.
• Should require little changes to standard
  practices.

RR DBR UBR
6
Data Collection

• How do we compute UBR?
• Treatment a set of customers who were contacted.
• Control a set of customer who were not
  contacted.
• similar to Treatment.
• All responders in Control must be decided buyers.
• UBR RR DBR

7
Model Construction
Characteristics exclusive to positive class
those of undecided customers.
Response
Yes
No
decided undecided
non
Treatment (T1)
(2)
(1)
Contact
non undecided
decided
Control (C1)
(3)
(4)
The learning matrix

• ltT1, C1gt training set,
• ltT2, C2gt validation set.

8
Proposed Solution Model Evaluation

• Rank ltT2, C2gt
• T2x top x of the ranked list of T2 (contacted),
• MT RR of T2x,
• C2x top x of the ranked list of C2 (not
  contacted),
• MC RR of C2x.
• T2x and C2x are similar,
• UBR RR DBR MT MC

9
Related Work Los

• Predict the amount of positive influence the
  contact has on each customer.
• Positive class responders,
• Negative class non-responders,
• Use treatment variable T to describe if a
  customer has responded. However,
• T 1 needs to be more strongly associated with
  the positive class.

similar to traditional paradigm
10
Experimental Evaluation
Response Response
Yes No
Treatment (1) 1,182 (2) 20,816
Control (3) 108 (4) 2,400

• Data real campaign for a loan product.
• 3-fold cross validation.

Our influential approach
Traditional paradigm
Los
Our influential approach oversample (3)
11
Joint Comparison

• Improvements of our approaches are significant in
  the top percentiles.

Association Rule Classifier