4. 4
• Very competitive environment: Competition, government,
economy, tight margins, 3G and mobile internet, and so on
• Customer Saturation = Skyrocket Churn
• Retention is the name of the game
• Quality over money (Content is the king!)
Introduction
Value-Added Services (VAS) industry
5. 5
• Ensemble Learning using Survival Analysis and Deep Learning
• Alter that we’ll score our subscribers
• Each score range will have a special treatment
• Development of retention strategies: Price elasticity, cross-
selling, special offers, channel acquisition analysis, up-front
payments, freemium strategies, promotional strategies, pay as
you go
Problem Definition – How to stop the Churn?
Prescriptive Churn Modeling
6. 6
•Survival function S(t):
T = Event time
f(t) = Density function
F(t) = Cumulative Density function.
Survival Analysis
Statistical procedure that estimates Time-to-Event probability of survival
8. 8
• Several processing hidden layers
• High complexity structures
• Non-Linear problems
• Sorry guys but it’s a very old idea (Ivakhnenko, A. G. (1971).
Polynomial theory of complex systems)
• Deeply applied in Computer Vision
Deep Learning
Abstraction, processing, no-linearity. Together.
11. 11
What worked?
• Ensemble Learning
• Pipeline of data integration (Use
Amazon Redshift as meat grinder)
• Weekly processing
• Statistical Sampling
• Fast experimental approach (No long
meetings, no bureaucracy, quick-win
modeling, etc.)
What did not worked?
• Daily processing
• Stand-alone solutions like Weka,
Statistica, Python
• Traditional approaches of Machine
Learning like SVM and Logistic
Regression
• H2o.ai distributed processing using R or
python
Conclusions and Recommendations
Warfield hard lessons