Post on 04-Jan-2016
transcript
How Data Science Helps Prevent Churn at Avira,a 100-million User Company
Calin-Andrei BurloiuBig Data Engineer
Iulia PașovMachine Learning
Engineer
Strata + Hadoop WorldNew York, 2015
About Avira
• Headquarters in Tettnang, Germany
• Security applications for– Windows– Mac OS– iOS– Android
• Awarded for malware detection
Big Data at Avira
• 430 million global installs• 100 million users• On-premise Hadoop cluster– 7 worker nodes– 30 TB logs and events– 5 TB monthly new data
About User Churn
Active Installs
New Installs Uninstalls
Steps
Diagnosis
What can we measure?
Which are the churn reasons?
Understanding
Why do users have issues?
Who is likely to churn?
Treatment & Prevention
How can we react to prevent this?
Churn DiagnosisWhat can we measure?
Which are the churn reasons?
What can we measure?
• Metrics– Churn rate– New Installs– Active users– Usage patterns
Computing Churn from Uninstall Events
• Uninstall events collected as application logs
• Pros:– An event is an uninstall for
sure• Some users reinstall
• Cons:– Some events are lostoffline
online
Computing Churn from User Inactivity
• Check user event logs• Users are considered churned after some time of inactivity• Pros:
– More accurate• Cons:
– Requires waiting– Results come too late
0 10 20 30 40 50
Days
user inactive for 30 daysuser returns in the 31st day
User Inactivity Convergence
1-Apr 11-Apr 21-Apr 1-May 11-May 21-May 31-May0
50
100
150
200
250
3118 10In
acti
ve u
sers
Estimating User Churn
• Predict monthly user churn rate– Predictor
• uninstall events– Outcome
• inactive users
Apr May Jun Jul Aug Sep0
20
40
60
80
100
uninstall inactivepredicted
Performing Survival Analysis
Jul-1
4
Aug-
14
Sep-
14
Oct-1
4
Nov-1
4
Dec-1
4
Jan-
15
Feb-
15
Mar
-15
Apr-1
5
May
-15
Jun-
15
Jul-1
5
Aug-
15
Sep-
150.0%
20.0%
40.0%
60.0%
80.0%
100.0%
60%
Su
rviv
al P
rob
ab
ilit
y
User Profile
• Consider– Devices– Behavior– Technical savviness– Business or consumer?– Errors
Users
User Profile
Churned Users
Active
Churned
Uninstall Surveys
• Ask users to complete a survey on uninstall
• Find churn reasons• 1% users complete surveys• Complaints from the past
Uninstall
Surveys
Lifecycle Surveys
• Complaints from the present• Ask users to give feedback a
few weeks after installation• Questions based on insights
from uninstall surveys
• Market research– Know your product’s
market
Lifecycle
Surveys
Extracting Sentiments from SurveysUninsta
ll Survey
s
Lifecycle
Surveys
Sentiment
Analysis
• Sentiment analysis– Negative review
• Dissatisfaction– Positive review
• Arbitrary reasons (e.g. reinstall)
Extracting Churn Reasons from Surveys
• Topic detection– Churn reasons
• Insights might be misleading
Uninstall
Surveys
Lifecycle
Surveys
Sentiment
Analysis
Topic Detectio
n
Reasons
Churn UnderstandingWhy do users have issues?
Who is likely to churn?
Matching Profiles with Reasons
• Compare users– With churn
reasons– Loyal
• Find patterns– Characteristics– Behavior– Context
Uninstall
Surveys
Lifecycle
Surveys
Sentiment
Analysis
Topic Detectio
n
Reasons
User Profile Match
How Avira Identified Churnable Users
• Uninstalled surveys revealed an “update” issue as a churn reason– “The product could not update so I uninstalled.”
• User profile of users with the “update” problem– Context
• A particular version of the antivirus– Behavior
• Antivirus didn’t update for at least 2 weeks• Users were active at least 4 times in 2 weeks
Churn Treatment & Prevention
How can we react to prevent this?
How can we help?• Find solutions for each churn reason• Directly
– Fix bugs– Fix UX– Add requested features– Offer the right price for extra features
• Indirectly– Head them to support team
To Summarize...• Know your data• Diagnose users who leave• Find and understand reasons• Treat every reason to prevent churn
Acknowledgements• Many thanks to our colleagues who worked with us on this project
or helped us with the presentation• Rodica Coderie
• Data Scientist
• Viacheslav Rodionov• Big Data Engineer
• Anna Tyrkich• Designer