- Host : Kaggle
- Partner : Kharagpur Data Analytics Group
- Duration of competition : May 3, 2021 -> May 11, 2021
- For more details visit competition page
-
Structure of Dataset is as follows:
- train.csv [Contains training data]
- test.csv [Contains 25% of test data]
- Sample Submission.csv [sample csv file for submission]
-
Prediction column :
Customer Attrition -
Unique ID of customer :
ID -
Dataset contains 14 features related to telecom data:
`Sex`, `Aged`, `Married`, `Total Dependents`, `Service Span`, `Mobile Service`, `4G Service`, `Cyber Protection`, `Hardware Support`, `Technical Assistance`, `Film Subscription`, `Settlement Process`, `Quarterly Payment`, `Grand Payment` -
Quarterly Payment,Grand Paymentare continuous data columns [In some experimentsService Spanis also considered as continuous] and remaining are Categorical columns.
- I used
CatBoost,XGBoost,LightGBM,Histogram Gradient Boostingtechniques for training. - After training based on obtaied feature importances, Various features were removed and experimented with different combination of features.
- In some cases I changed
Service Spancolumns as continuous orBucketed data, but there is no much change in leaderboard score. - Finally the best models are ensembled and achieved 80.79% accuracy on 2800 unseen data samples.
- Even though I experiemnted with many models and enselbling models, kfold cross validation, leaderboard score saturated at around 81% accuracy.
- Following is one of the feature importance plot obtained from CatBoost training. [Use light model for better visualization of image :)]
- models folder contains the sample models (
.dumpformat) obtained from 5 fold cross validation training with CatBoost model.
- This folder consists of
logsobtained during training ofCatBoost Classifier.
- The order of score with various models given same data pipeline is as follows : Cat Boost > Light GBM > XG Boost > Histogram Gradient Boosting
- Note : This order varies with varying Training data and features.