BigData Republic teamed up with VodafoneZiggo and hosted an meetup on churn prediction.
Telecom companies like VodafoneZiggo have long benefited from the fine art/science of predicting churn. Currently, in the booming age of subscription based business models (e.g. Netflix, Spotify, HelloFresh), the importance of predicting churn has become widespread. During this event, VodafoneZiggo shared some of its wisdom with the public, after which BDR Data Scientist Tom de Ruijter presented an overview of the modeling tools at hand, both classical, as well as novel approaches. Finally, the participants engaged in a hands-on session showcasing the implementation of different approaches.
PART 1 — Churn Prediction in Practice by Florian Maas
At VodafoneZiggo we are incredibly excited about Advanced Analytics and the enormous potential for progress and innovation. In our state of the art open source platform we store the tremendous amount of data that is generated every single second in our mobile and fixed networks. This means that we have a vast body of rich information, which if unlocked, can lead to something very special. As a company with a primarily subscription-based service model, churn plays a vital role in the daily business. Not only is the churn rate a good indicator of customer (dis)satisfaction, it is also one out of two factors that determines the steady-state level of active customers. During this talk, we will show how data science provides added value in the process of churn prevention at VodafoneZiggo. We will talk about the data and the modeling approach we use, and the pitfalls and shortcomings that we have encountered while building the model. We will also briefly discuss potential improvements to the current approach, which brings us to talk #2.
PART 2 — The Churn Prediction Toolbox by Tom de Ruijter
The second talk will show you the fine intricacies of predicting churn through different approaches. We’ll start off with an overview of different modeling strategies for describing the problem of churn, both in terms of a classification problem as well as a regression problem. Secondly, Tom will give you insights in how you evaluate a churn model in a way such that business stakeholders know how to act upon the model results. Finally, we’ll work towards the hands-on session demonstrating different model approaches for churn prediction, ranging from classical time series prediction to recurrent neural networks.
7. 8
• This presentation:
- The current approach to churn prediction within VZ
- Lessons learned & challenges faced while building this model
• Second presentation:
- More complex modelling approaches
TODAY
8. 9
• Florian Maas
• Econometrics & Management Science, specialization in
Operations Research and Quantitative Logistics
• Supply Chain Improvement Specialist at Interface
• Co-founder of Xaperi, started joint-venture with Cadran.
• Data Scientist at VodafoneZiggo since jan 1st.
• I like running, hiking, guitar, concerts, field hockey, cycling,
boxing(?), Calvin & Hobbes and capybara’s.
WHO AM I?
9. 10
VodafoneZiggo at a glance
#1 in Cable
8,000 employees
#2 in Mobile
4m TV connections
2.5m fixed phone connections
3m broadband internet connections
5m mobile voice/internet connections
€4b revenue
Cable Network
Coverage
Mobile
Network
Coverage
7m homes passed
VodafoneZiggo Kick off
10. 11
A Dutch company with a global scale
• €55.9bn Revenue
• 108,000 employees
• 523m mobile customers
• €20bn Revenue
• 45,000 employees
• 50.1m homes passed
VodafoneZiggo Kick off
11. 12
What we do really matters
For our customersFor our people For society
VodafoneZiggo Introduction
12. 13
NIELSEN’S LAW OF INTERNET BANDWIDTH
Summary: Users' bandwidth grows by 50% per
year (10% pts. less than Moore's Law for
computer speed).
13. 14
WHO ARE WE | ADVANCED ANALYTICS
Our mission is to unlock the value hidden in our huge pile of
data and translate it into valuable insights and data driven
products for our customers, business and stakeholders using
techniques from the field of Machine Learning and Artificial
Intelligence
15. 16
WHAT IS CHURN?
Churn rate (sometimes called attrition rate), in its broadest sense, is a measure of the number of individuals or items
moving out of a collective group over a specific period.
16. 17
HOW CAN ADVANCED ANALYTICS HELP PREVENT CHURN?
Provide a better service to our customers.
• Network optimization
• Where to place new broadcasting
equipment?
• Proactive monitoring of the network.
Provide special offers to customers.
• Who is likely to churn?
• Which offer should we give to which
customer?
Understanding the churn drivers
• What drives a customer to churn?
17. 18
• Team:
• Platform:
• Data:
• Goal: Get a working model up and running quickly and pave the way for more complex modeling
approaches.
INITIAL APPROACH
19. 20
B2C FIXED CHURN PROPENSITY MODEL
All Ziggo customers Machine Learning model
0.92
0.85
0.63
0.62
0.59
0.46
0.32
0.02
Ziggo contacts customers with top
churn predictions on weekly basis
with service call
Monthly output of
customer predicted
propensity to churn
…
24. 25
DATA USAGE BOARD (DUB) & GDPR
Data of customers that have left
us some time ago
Network traffic
Certain TV package
subscriptions, such as erotic or
Arabic.
Customer data (gender, age,
customer lifetime, …)
Active products
Household information bought
from external parties
Limited data on interactions
with the help desk
26. 27
PRODUCT DUMMIES – INFORMATION LEAK IN MODEL?
• This variable will be considered very important by the model and the model will draw incorrect conclusions based
on this variable.
Insert presentation title via header & footer
Product
introduction
Jan 1st April 1st
May 1st
Product unavailable; dummy 0. Product available; dummy 0 or 1
27. 28
PRODUCT DUMMIES
Insert presentation title via header & footer
Current solution:
• Only include a dummy if a product was available in the entire period.
today – one year today – ½ year today
Included in dataset
Included in dataset
Excluded from dataset
Excluded from dataset
29. 30
CUSTOMER INTERACTIONS: FUTURE
Sentiment analysis on interactions:
• Sounds nice, but this has not passed the DUB yet.
• "Calls may be recorded for training and quality purposes“
31. 32
DATASET
Customer id Feature 1 Feature 2 … Churned?
1 x 0.2 0
2 x 1 0
3 y 0.5 0
4 y 0.3 1
5 z 0.2 1
6 z 0.1 0
Customer id Feature 1 Feature 2 … Churned?
1 x 0.2 0
2 x 1 0
Customer id Feature 1 Feature 2 … Churned?
3 y 0.5 0
4 y 0.3 1
Customer id Feature 1 Feature 2 … Churned?
5 z 0.2 1
6 z 0.1 0
Customer id Feature 1 Feature 2 … Churned?
1 x 0.2 0
2 x 1 0
3 y 0.5 0
4 y 0.3 1
Customer id Feature 1 Feature 2 … Churned?
5 z 0.2 1
6 z 0.1 0
Customer id Prediction
5 0.05
6 0.25
32. 33
Customer id Feature 1 Lifetime Months in
future
6 b 10 0
6 b 11 1
6 b 12 2
Customer id Feature 1 Lifetime
6 b 10
Customer id Feature 1 Lifetime Months in
future
Prediction
6 b 10 0 0.0
6 b 11 1 0.0
6 b 12 2 0.7
Customer id Prediction
6 0.7
34. 35
MORE IMPROVEMENTS FOR THE FUTURE
• Better predictors: CSQ, 2020 errors, …
Predict which retention offer to give to which customer.
Make better use of the temporal component in the modelling of this problem.
1
3
4
Predict which customers to target, rather than predicting which customers are most likely to
churn.
2
45. Agenda
1. Churn: what, why and how?
2. Machine learning for churn prediction
3. Use case: churn classification at VodafoneZiggo
46. Capturing customer life cycle into features
time
CHURN
Prediction window
Calls for time-series prediction
PROBLEM: we know very little about customers
47. Capturing customer life cycle into features
PULL PUSH
Competing products
Campaign history
Market presence
Spending behavior
Customer loyalty
Client interaction
Product usage
Device logs
Event logs
Customer loyalty
48. System design
Churn prediction algorithmcustomer data
!(#)
output
static time-series Churn in x days? Probability on time
49. Churn prediction algorithm
!(#)
Business rules simple, explainable > baseline
Classification model
churn in the next N days? Y/N
pre-defined window size
Probability estimation
can use censored data
flexible predictions
50. Client data Model Output
∑ > 3?
Baseline: business rules
=activity
=inactivity
Take action
51. A simple linear model
one-hot
encoded
categorical
features
…
output
!"#
$#%numerical
features
&#%