Credit Cards Customers Project

Project Description

Customer Churn (customer attrition) is the most challenge problem for business such as credit cards or telecommunication companies etc. Building models to predict who is going to churn would help improve business, and companies can prevent from losing their customers.

In this project, I analyzed credit card customers' dataset and built machine learning (ML) models to predict who churn the service. The roc-auc score of the final model is 0.993.

More details for the dataset is here. I also wrote a blog for this project and posted in Kaggle.

Machine Learning Model Pipeline

I compared 6 models (Logistic Regression, Support Vector Machine, KNeighbor, Random Forest (RF), AdaBoost (ADA), and GradientBoostingModel (GBM)) by using results' metrics (Accuracy, Recall, Precision, and ROC AUC score).

GBM model showed the best result compared to other models.

Tuned GBM model shows better performance compared to the baseline GBM model.

By using the best hyperparamters, we got an improved GBM model showing 7% increased performance (0.993 of ROC-AUC score) compared to the baseline (0.916 of ROC-AUC).

NOTE: Please see Building_Model_Pipeline.ipynb, if you want to see more detail.

Dealing with Imbalanced dataset

We have a classification problem. We have 84% of Existing customer data and 16% of Attrited customer data. This dataset is imbalanced. So, we needed to deal with this very carefully.

I compared 5 different over-sampling methods: RandomOverSampler, SMOTE, ADASYN, BorderlineSMOTE, and SVMSMOTE.

As a result, ADASYN and BorderlineSMOTE showed the best performances so that I chose ADASYN to balance the dataset in the final pipeline.

NOTE: Please see Oversampling.ipynb, if you want to see more detail.

Visualization_EDA

Creating graphs help solve the questions below:

1. Any differences between Churn and Exist groups depending on each feature? Some features show differences between Churn and Exist groups.

2. Are there correlations between numerical variables?

NOTE: Please see Cleaning_EDA_and_Visualization.ipynb, if you want to see more detail.

File Description

The files structure is arranged as below:

- Data
    - BankChurners.csv: raw data 
    
- Building_Model_Pipeline.ipynb 
    : Workflow regarding building Gradient Boosting Model.
- Cleaning_EDA_and_Visalization.ipynb 
    : Data cleaning, exploratory data analysis, and visualizing data.
- FeatureEngineeringScaling.ipynb 
    : Investigations regarding feature engineering, feature scaling, and feature importance. 
- HyperparameterTuningForGBM.ipynb 
    : Hyperparameter tuning by using Optuna to find the best parameters for GBM model. 
- Oversampling.ipynb
    : Finding the best over-sampling methods for this project to deal with imbalanced dataset.
- readme.md

Dependencies

• Python 3.5+
• Machine Learning Libraries:
  • Numpy
  • Pandas
  • Sciki-Learn
  • Feature-engine
  • Imblearn
  • Optuna
• Visualization: Matplotlib, Seaborn

Acknowledgements

Data was provided by Kaggle