1. From Knowledge Discovery to Customer Attrition
CCI, UNC-Charlotte
From Knowledge Discovery to Customer Attrition
Katarzyna Tarnowska Zbigniew W. Ras
San Jose State University Univ of North Carolina
Dept. of Computer Science Dept. of Comp. Science
San Jose, CA 95192, USA Charlotte, NC 28223, USA
Research sponsored by
ISMIS 2018, CYPRUS

2. CLIRS - Customer Loyalty Improvement Recommender System
(for improving business revenue)
Research sponsored by

3. The Daniel Group - Consulting Company
What is the plan
Build
Recommender System
for each client
(34 clients) helping to
increase its revenue
Client 1
Client 2
Client 3
Client 4
shops
shops
shops
Build Personalized
Recommender System
for each shop helping to
increase its revenue
Services
(heavy equipment repair)
Parts
CUSTOMERS

4. NPS based approach Net Promoter Score (NPS) –
today’s standard for measuring customer loyalty
Promoter - {9,10}
Passive – {8,7}
Detractor – {1,2,3…,6}

5. Get-Keep-Grow model based approach

6. NPS rating for all clients

7. NPS rating for all shops

8. CLIRS
Our System

9. Example of benchmark questions:
Benchmark All Overall Satisfaction
Benchmark All Likelihood to be Repeat Customer
Benchmark All Dealer Communication
Benchmark Service Repair Completed Correctly
Benchmark Referral Behavior
Benchmark Service Final Invoice Matched Expectations
Benchmark Ease of Contact
Benchmark All Does Customer have Future Needs
Benchmark Service Tech Promised in Expected Timeframe
Benchmark Service Repair Completed When Promised
Benchmark Service Timeliness of Invoice
Benchmark Service Appointment Availability

10. Decision Table built from customers survey
300,000 records per year,
Customer
M/F
Phone
Bench1
Bench2
Bench3
Comments
Promoter Status
2011 M 3 9 5
text1
Promoter
2012 8 10
text2
2013 F
text3
Detractor
2014 7
text4
Passive
New attributes can be derived
Text Mining & Sentiment Mining
can be used to build new attributes

13. Extracting Meta-Actions
DATASET
Guided Folksonomy &
Sentiment Analysis
Meta-Actions
Extracting Comments C1 C2 C3 C4
C2,C4
Knowledgeable
Staff
Customer
Bench1
Bench2
Bench3
Comments
Prom_Stat
Cust1
high
med
high
C1, C2
Prom
C1,C3
Friendly Staff
Cust2
med
med
med
C2, C3
Pass
C1 refers to Bench2
C3 refers to Bench1
C2,C4 refers to Bench3
Cust3
C4, C3
Benchmark Values ={low, medium, high}
Extracted Rules (example)
R1=[ (Bench1=high)&(Bench2=med) …………  (Prom_Stat = Prom)] sup(R1)={Cust1,..}; Let’s say Confidence=90%
R2=[ (Bench1=med)& ……………. (Bench3=med)  (Prom_Stat = Pass)] sup(R2)={Cust2,..}; Let’s say Confidence=95%
Action Rules
R= [(Bench1, med ->high)&(Bench2=med)  (Prom_Stat, Pass -> Prom)] Sup(R)={Cust2,..}; Confidence=90%*95%=85.5%.
Recommendation: Staff has to be more friendly (R will be activated)

16. CLIRS - Customer Loyalty Improvement Recommender System
1. NPS Module based on benchmarks
2. NPS Module based on text data
3. Customer Attrition Module

17. 2. NPS Module based on text data
Using Folksonomy & Sentiment Analysis to Find New Features

18. Meta Attributes
Service
Example of action rule: (Tag2: -1 0)  (Prom Stat, Detractor  Passive)

19. 3. Customer Attrition Module
Data Available:
Survey Data (customer activity) and customer feedback for all clients
have been collected for the years
Sales (transactional) Data on customers for all 34 companies, years
Our Approach:
Determine if there are markers in the sales trends (Sales Data) that might suggest
a customer is getting ready to defect.
Mine for patterns in the Survey Data that would suggest likelihood to defect.
Compare results with the actual data on the defected customers.

21. Automatic 2-category data labelling algorithm
Label surveyed customers as “Leaving – No” when:
There are transactional records in next years for that customers
Label surveyed customers as “Leaving – Yes” when:
There are no transactional records for that customer in the following years
Label is unknown if:
The Invoice count for the current year is unknown (NULL)
Data Transformation
- Extract Client22 customer data from years
- Add column denoting „Invoice count” based on additional transactional data
- Add column „Leaving” and apply labelling algorithm on data based on transactional data
Example
Generate Action Rules
IF (Benchmark1, 3 6) AND (Benchmark2, 7 9) THEN (”Leaving-yes”  “Leaving-no”)
No transaction data was found for that customer after 2014, therefore it was labelled as defected

23. Automatic 3-category data labelling
“Active” customers – are these actively making transactions each subsequent
year after they were surveyed.
Growing (in terms of no. of transactions)
Declining (in terms of no. of transactions)
“Leaving” - customers who might be at risk of losing them,
as the transactional data showed that they stopped doing business for one year
(no transactional records in a year following the survey).
“Lost (defected)” - customers who stopped doing business continuously
for two years.
Note: 2016 customers can be only „Active” or „Leaving” (since we have data up to 2017)

25. Example of Classification Rule extracted from the dataset:
(Benchmark : Service − Final Invoice Matched Expectations  1) and
( Likelihood to Repurchase  6) and ( Likelihood to Repurchase  4) and
(Benchmark : All Dealer Communication  4) and
(Benchmark : All Ease of Contact  6) => CustomerStatus=Lost; Sup=16, Conf=1
Example of Action Rules extracted from the dataset:
“Survey Type” was selected as a stable attribute, and all the “Benchmark” attributes
were chosen as flexible attributes
SurveyType(Field):
(BenchmarkAllEaseofContact (10)  BenchmarkAllEaseofContact (8)) 
(CustomerStatus(Active)  CustomerStatus(Leaving)), Conf = 0.82
(BenchmarkAllDealerCommunication (10)  BenchmarkAllDealerCommunication (8))  (CustomerStatus(Active)  CustomerStatus(Leaving)), Conf = 0.82
SurveyType(Field): (BenchmarkServiceRepairCompletedCorrectly (10)  BenchmarkServiceRepairCompletedCorrectly (9)) 
(CustomerStatus(Active)  CustomerStatus(Leaving)), Conf = 0.81

26. More Action Rules Extracted:
Survey_Type(Field) : (Benchmark Service-Repair Completed Correctly (10  9) 
CustomerStatus(Active  Leaving) Conf = 0.81
Survey_Type(Field) : (Benchmark Service-Repair Completed-When-Promised(10 9 ) 
CustomerStatus(Active  Leaving) Conf = 0.8
Survey_Type(Field) : (Benchmark Service-Repair Completed-When-Promised(8 10) 
CustomerStatus(Leaving  Active) Conf=0.8
Survey_Type(Field) : (Benchmark All-Overall-Satisfaction(8 9) 
CustomerStatus(Leaving  Active) Conf = 0.76
Survey_Type(Field) : (Benchmark Dealer-Promoter-Score(8 10) 
CustomerStatus(Leaving Active) Conf = 0.77
NEXT: Mine for action rules triggers (meta-actions)