Social Networking for Churn
Analysis
Predictive Analytics World 2011 SF
David Katz
Dataspora
David Katz Consulting
Outline of Case Study
•
•
•
•
The Project
The Goals
Methods
Findings
The Project
• Mobile Phone Provider with concern about losing
subscribers
• Uses Predictive Analysis (Logistic Regression) to
target subscribers with increased risk of churn
• Looking for ways to improve prediction and
retention
• Pilot Project identified “Social Networking” type
data
Goals
• Can Social Networking improve the existing
churn model?
• This is a means to an end – The ultimate goal
is to reduce churn.
Reducing Churn
• 1) Identify a group of subscribers with high
risk of unsubscribe.
• 2) Target effective intervention to this group without offering expensive incentives to many
who will not unsubscribe.
• This can be a high bar!
Predicting Unsubscribe Rate Can Help
Target Cost-Effective Interventions
Churn Rate
70%
60%
50%
40%
30%
20%
10%
0%
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
Preliminary Univariate Analysis
• Hypothesis – subscribers who have been
calling unsubscribers are more likely to
unsubscribe.
• Calling network – who is calling whom?
Social Network Analysis
More Connections found between
Cancellers than expected by chance
April 2009
21 customers
Dynamics of cancellation in a
selected customer call network
May 2009
2 deactivations
Dynamics of cancellation in a
selected customer call network
June 2009
4 deactivations
Dynamics of cancellation in a
selected customer call network
July 2009
7 deactivations
Dynamics of cancellation in a
selected customer call network
Cancellation Rates Differ
for customers calling cancellers
Subscriptions at Start of June
Cancellations among this group in June
Cancellation Rate
Subset with known calls to May churners
Subset of these churning in June
Cancellation Rate
1980582
26294
1.3%
35387
837
2.4%
And this persists over time…
3.0%
2.5%
2.0%
Overall Cohort
1.5%
Called Cancellers During May
1.0%
0.5%
0.0%
June
July
August
Social Networking Data
• Association
– Relatively Stable
– May be correlated with other predictors
– How much can it add to existing churn model?
• Influence
– Related in time
– May be transient
Variable Selection
• Initial variable selection done using “earth”,
the R implementation of Multivariate
Adaptive Regression Splines
• Can recognize nonlinear predictors effectively
• Extremely fast algorithm reduces computation
by computing the cross products
incrementally
Social Network Variables
• Inbound callers “closer” than outbound
callees.
• Voice calls are much better measures of
association than SMS/texting.
• Counting Calls slightly better than overall
length of calls.
Main Social Networking Variable
Selected
• Percentage of incoming voice calls in the last
90 days which were from customers cancelling
in the last 30 days.
• Those with 10% or higher on this variable
were designated as churn-chatters – those
speaking with recent unsubscribers relatively
frequently.
Generalized Additive Models
• Generalization of Generalized Linear Models
• Using family=“binomial” makes it a
generalization of Logistic Regression
• Allows nonlinear predictors (without
transformations or binning)
Generalized Additive Models
• Used mgcv::gam – the Implementation in R
• Logit(p) = ∑ f(x)
• Automatic plotting of each smooth f(x) term
Smooth Term Example
Days to Contract End
Timeline
• Timing is an essential dimension
– When can we obtain the data?
– When will the intervention impact the customer?
– What time window should we use for evaluating
the cancellation rate?
• Prescribed by the client:
– 60 day lead time
– 60 day window for measuring resulting churn.
The baseline model predicts seems to predict
churn reasonably well when we look at the
complete set of subscribers…
0.06
0.04
0.02
Fraction
Churn
Actual
Pct
Churn
Actual
0.08
Fit 60-120 day churn with no social network data
0.02
0.04
0.06
Predicted w/o Social Networking
Each dot represents one semidecile of the subscribers
0.08
…but misses the systematic bias when we plot the
churn-chatters.
0.15
Subset with
> day
10%period
30 days of calls predicting
churnchurn.chat
over a 60-120
0.10
Predicted w/o
social networking
0.05
Fraction
Churn
Actual
Pct
Churn
Actual
Actual
0.02
0.04
0.06
0.08
0.10
0.12
0.14
Predicted w/o Social Networking
* Churn-chatters are those who have more than 10% incoming calls coming from churners
The impact of social networking on churn-chatters is
larger if we look at the churn window for the
immediate next 30 days
0.15
0.20
Actual
0.05
0.10
Predicted w/o
social networking
0.00
Fraction
ChurnChurn
ActualActual
Pct
0.25
0.30
30 days of calls predicting next 30 days of churn
0.00
0.05
0.10
Predicted w/o Social Networking
0.15
The 4-month window gives us the most comprehensive look at
churn-chatters
0.4
30 days of calls predicting next 120 days of churn
0.3
0.1
0.2
Predicted w/o
social networking
0.0
Fraction
ChurnChurn
Actual
Pct
Actual
Actual
0.05
0.10
0.15
0.20
Predicted w/o Social Networking
0.25
0.30
There is great value in reacting faster as churn-rate
declines rapidly over time – especially in high risk
groups identified by model and churn chat
0.15
0.05
0.10
Feb
Mar
Apr
May
0.00
Pct Churning in Each 30-day period
Comparing Churn Rates Over Time
All EM/EM+ Subs
Top 5% Per Base Model
No Social Networking in the Base Model
Top Churn Chatters
Percentage impact on churn propensity
Impact of Social Networking Highly
Dependent on Time to end of contract
Days to Contract End
Red – churn chatters
Black – all others
Not a Proportional Hazard
• The Social Networking effect is markedly
different depending on “Days to Contract End”
• The greatest SN effect is closer to the actual
end of contract.
• Additional SN variables are
also in the model
Lessons Learned
• Those receiving calls recently are less likely to churn.
Lessons Learned
• Social Networking is composed of at least two effects –
• Shows affiliation
• In our model this is an indicator variable with a relatively small value,
constant over time.
• Has influence
• Immediate influence decays over time
• Higher at the critical time near end of contract
Lessons Learned
• Adding social networking variables does help identify more churners
• Especially in the high-risk group of churn-chatters: those who get more
than 10% of the calls in the last 90 days from people who have churned
in the last 30 days.
0.3
0.2
0.1
0.0
Actual Churn Pct
0.4
30 days of calls predicting next 120 days of churn
0.05
0.10
0.15
0.20
Predicted w/o Social Networking
0.25
0.30
Lessons Learned
• Time is of the essence!
• Experimenting beyond the box of current timing constraints reveals
new features in the data.
• SN has the strongest immediate effect (days) and then decays
over a longer period (several months)
Time since Caller Churn
Regardless of Association Measure
Lessons Learned
• Time is of the essence!
• Interventions must happen right away or many customers will
already be gone.
0.15
0.05
0.10
Feb
Mar
Apr
May
0.00
Pct Churning in Each 30-day period
Comparing Churn Rates Over Time
All EM/EM+ Subs
Top 5% Per Base Model
No Social Networking in the Base Model
Top Churn Chatters