Analiza migracji klientów: jak przeprowadzić prognozę churnu z użyciem Machine Learning?

Wiktor Zdzienicki

1. Why is churn modeling important?

2. Challenges

3. Featured solutions

4. Project stages

5. Solution example – Azure ML Studio

6. RFM & Score

7. Solution example – Power BI

Monitor your business

It’s a crucial metric for a growing number of companies in a variety of industries

Traditional consumer behavior patterns disappear, along with the decrease of customer loyalty and retention

Companies must make use of the data to analyze not only the probability of churning but also combine this with an evaluation of customer value

Customer churn – why is this important?

New customer habits

Appreciate your existing customers

Not only churning

Acquisition of a new customer is much higher than the cost of retaining current clients

Challenges

• How to use analytics to determine the probability of customer churn?

• How to combine it with scoring and estimation of customer value?

Featured solutions

Churn

Churn analysis

Identification of the reasons which made a

customer leave

The final effect is the probability of situation in

which a customer stops using our products or

services

Benefits

o Lower cost of generating campaigns

o Specifying the negative effects of potential

churn

o Tailor-made and targeted marketing actions

Project Stages

Project Stages

Project Stages

• Removing outliers• Imputing empty records• Format correction• Binning & grouping• Target creation

• Enriching with dictionaries• Extracting dates• Interactions• Aggregations• Dummy features

Project Stages

• Exploratory data analysis• Correlations• Distributions • Log transformations• ANOVA

• Manual selection• Stepwise selection• LASSO• Feature importance

Project Stages

• Dividing data for test & train sets

• Complexity & interpretability discussions

• Selection of few rival models, eg.:• Logistic

regression• Random forest• Neural net• Gradient

boosted decision trees

Project Stages

• Generating predictions on known test data• Used for

investigating accuracy

• Generating predictions in production process• Used for making

business decisions

Project Stages

• Setting desired accuracy metric

• Analysingconfusion matrix in terms of:• Optimal

threshold• Accuracy• Recall• Precision

• Selecting the best model basing on selected metrics

Project Stages

• Models with default parameters do not give optimal results

• Tweaking range of input values in order to achieve the pest performance

• Optimisingcalculation time & model robustness

• Producing final results

Solution example

Modeling in Azure ML Studio