Team w: Andrea Huscher, Gaetano Costa, Federico Porcu
churn analysis for Data Mining proj module 2 Matteo Calia The teacher provides us unstructured data, so we were supposed to clean it, visualize it and then process it. The aim was also to provide economic insights to the 'management' so that the statistical models and all the visualization works were not for their own sake, indeed oriented to a target and ready-to-go management decision.
- EDA part:
- explorative distribution histograms, bar plots
- interactive Plotly
- 3D data analysis, multivariate data visualization
- economic implications through visualization
- Feature Engineering:
- fill missing obs and NAs, imputer, binned means (TotalCharges and Tenure)
- generate offer packages combination to cluster offers (reverse engineering)
- clustering Tenure both economic intuition (1 y rule of thumb) and kmeans algo
- compute the variable unnkown costs (given the inequality between monthlycharges * tenure != totalcharges )
- Variable Importance:
- Correlation matrix
- HeatMap w\
Seaborn
- Encoding:
- OneHot
- MeanEnc
- Train Test Splitting:
- sss
StratifiedShuffleSplit
- k fold cross
- upsampling to solve unbalance
- sss
- Models and hyper tuning:
- Random Forest
- XGBOOST
- Gradient Boosting
- Logistic l1 penalty
- Logistic l2 penalty
- SVM
- Radar plt to compare performance:
- Summarize results:
- LIME interpretation
- what data suggest to CMO