Churn Prediction with Predictive Analytics and Social Networks in R/Python

📅 May 23rd, 2019, 9am-4.30pm
🌍 English

Introduction

This comprehensive advanced course to analytical churn prediction provides a targeted training guide for marketing professionals looking to kick-off, perfect or validate their churn prediction models.

Combining theory with practice, this course starts with a brief recap on churn prediction and analytical modeling. We then elaborate on data preprocessing and discuss state of the art methods for feature engineering. Predictive analytics is covered next with various (recently developed) techniques discussed into sufficient detail. We also extend on social network analytics and illustrate how networks between customers can be leveraged and featurized to boost the performance of your analytical churn models. The course concludes by zooming out and discussing post processing of analytical churn models.

Throughout the course, the instructor(s) extensively report on their recent scientific findings and international consulting experience. The course provides a sound mix of both theoretical and technical insights, as well as practical implementation details. These are illustrated by several real-life case studies, reference models, state of the art research insights and benchmarks. We also use an example churn data set and provide both R and Python code for the various concepts discussed.

Course Outline

Introduction to Churn Prediction

what is churn and why do we need to predict it?
defining churn: contractual versus non-contractual settings
types of churn: passive, active, expected and forced churn
examples of churn in telco, banking, retail, FMCG and insurance
key characteristics of successful churn prediction models (model performance, model interpretability, model efficiency, model compliance)
churn prediction challenges: churn detection versus churn prevention, data availability, privacy and GDPR
analytical approaches to churn prediction
zooming out: Customer Lifetime Value (CLV) modeling and Uplift modeling
Introducing our example churn data set

Data Preprocessing for Churn Prediction

motivation (GIGO)
types of data (transactional, RFM, social media data, API/web scraping data, open data, clickstream data)
types of variables (continuous, categorical, nominal, ordinal)
denormalizing data
sampling and impact of sample bias
Example R/Python code: drawing a sample
visual data exploration (visual analytics, OLAP)
missing values
Example R/Python code: treating missing values
outlier detection and treatment (histograms, box plots, z-scores, Mahalanobis distance, winsorisation)
Example R/Python code: capping outliers
standardizing data
feature engineering: Box Cox, Yeo-Johnson, PCA, trend features, ratio features and continuity correction
Example R/Python code: Box Cox transformation
categorisation (pivot tables, odds analysis, Chi-squared analysis)
recoding categorical variables (dummy coding, Weights of Evidence, thermometer coding)
Example R/Python codee: WOE analysis
segmentation (statistical versus operational constraints)
variable selection (variable filters, Information Value, Wald’s Chi Square, Mutual Information)
Example R/Python code: computing the information value

Predictive Analytics for Churn Prediction

target definition and noise impact
linear regression
logistic regression (scorecards, nomograms)
Example R/Python code: estimating a logistic regression model
Example R/Python code: representing a logistic regression model using nomograms
decision trees (splitting, stopping, assignment decision)
ensemble methods: bagging, boosting, random forests, XGBoost
Example R/Python codee: estimating decision trees and ensemble methods
neural networks and deep learning
Example R/Python codee: building a neural network
dealing with skewed class distributions (undersampling, oversampling, SMOTE, ROSE)
Example R/Python code: dealing with skewed class distributions
evaluating churn prediction models: AUC, Gini, lift and Expected Maximum Profit (EMP)
Example R/Python codee: evaluating a churn model

Social Network Analytics for Churn Prediction

social networks and applications
explicit versus implicit (pseudo-) social networks
is churn a social phenomenon?
Statistical tests for homophily, dyadicity and heterophilicity
Example R/Python code: statistical tests for homophily, dyadicity and heterophilicity
social network components: vertices versus (weighted) edges
visualizing social networks (sociograms, adjacency matrices and lists)
social network metrics (degree, geodesic, closeness, betweennness, transitivity)
Example R/Python code: computing social network metrics
community mining (Girvan-Newmann, Q-modularity)
Feature engineering for social networks (link-based features, second-order link based features, neighborhood based features)
DeepWalk, node2vec and GraphSage
Propagating churn influence: Google PageRank, Personalised Google PageRank and Collective Inference Procedures
Example R/Python code: computing Google PageRank scores
Case study: Operationalising Social Network Analytics for Churn Prediction