KKBOX

KKBOX project on Kaggle

For details about the competition visit: https://www.kaggle.com/c/kkbox-churn-prediction-challenge/data

Project Overview

Objective: The main idea of the project is to predict churn for KKBOX customers.

Data:

• Transactions: Dataset with aprox 21.5 million records and features related with transactions of each customer: transaction date, end date, payment method, etc. Amount of unique records: 2.3 million records
• Users: 6.7 million unique records with information about the customer (age, city, gender, etc)
• Logs: 30 million records with customer behavior per day (# of unique songs played, total seconds played, etc). Amount of unique records: 1.8 million records
• Train: Aprox 1 million records indicating if customer has churn or renewal.

Code and Resources used:

Python Version: 3.7

Packages: pandas, numpy, sklearn, matplotlib, seaborn, streamlit

Fixing Oversampling: https://www.youtube.com/watch?v=OJedgzdipC0

For every dataset I followed this strategy:

1. Exploratoy Data Analysis (EDA)

• Loading data
• Statistics of the data frame
• Data visualization
• Data Cleaning
• Fixing formats

2. Feature Engineering

There are 2 challenging datasets (logs and transactions) because they have duplicated records due to their nature.

For logs dataset I followed these steps:

• Create features based on the average of the original feature.
• Data transformations: Most of the features were skewed to the right. I used the log transformation.
• Outlier detection

For transactions I followed these steps:

• Create features: Group the original dataset and for each customer create:
  • Amount of transactions.
  • Most frequent plan.
  • Most frequent payment method.
  • Revenue (sum of the amount paid for each transaction)
  • Auto renewal (most frequent decision)
  • Amount of cancelations (sum of cancelations for every user)
  • Most frequent quarter of transaction (the most frequent quarter of the year when the customer makes transactions)
• Outlier detection

After dealing with all datasets separately, I merged all datasets in one and then I applied the strategy of Exploratory Data Analysis and Feature Engineering again.

Customers with 0 and 2 years of tenure are more likely to have churn.

3. Modeling and Evaluation

The data has a problem bacause it's unbalanced. Almost 94% of data has no churn Vs 6% with churn.To solve this I used oversampling to balance data. The resource of how to use oversampling is the Resources section of this Repo.

Baseline Model:

Naive Bayes Classifier

ROC AUC Score	Accuracy	F1 Score
79.68	79.04	28.95

I will use the ROC AUC score as final metric to evaluate model performance..

Robust Models

I used robust models to predict churn: Decision Tree, XGBoost, and Random Forest. Here are the results for each model.

	ROC AUC CV	ROC AUC Test
Decision Tree	97.41	69.22
Xgboost	92.65	92.77
Random Forest	99.98	94.01

Random Forest is the best model to predict customer churn with a really good ROC AUC score.

Feature Importance

From the above, the features that have most predictive power are: regist_cancels is_auto_renew regist_trans revenue tenure

Deployment

For this project I used Streamlit to deploy a Web App