Customer Churn Classifier

A multi-model supervised classification workflow predicting telecom customer churn, built in KNIME and in Python

KNIME Workflow

Overview

An end-to-end classification pipeline that goes from EDA through preprocessing and tuning to a soft-voting ensemble
KNN and SVM are each built twice, once with native KNIME learner nodes and once in Python, and all five models are compared on the same test set
The data is imbalanced (always guessing "No" already scores 73% accuracy), so models are ranked by macro-F1 and Cohen's kappa rather than raw accuracy
EDA-driven feature selection. Ranked categorical features by the standard deviation of churn rate across their classes, and compared numeric distributions with churn-conditional box plots. Features with little separation (e.g. Gender) were dropped.
Removed redundant categories. Found that six service features share an identical "No Internet Service" class that duplicates the Internet Service feature, and merged it into "No".
Justified preprocessing. Stratified train/test split, imputed 0 for the missing Total Charges values (all new customers with tenure = 0), z-score normalization, and one-hot encoding with the redundant No_* columns dropped.
Tuned KNN and SVM on both platforms. Raising k from 3 to 17 lifted kappa from 0.34 to about 0.46. The SVM uses an RBF kernel with low C, and Python's gamma was converted to KNIME's sigma parameter.
Soft-voting ensemble of the tuned KNN and SVM plus Random Forest. It scored the best macro-F1 (0.737) and kappa (0.475) of all five models.

Layer	Tools	Description
EDA	KNIME (GroupBy, Pivot, Box Plot, Pie Chart)	Churn-rate aggregation per class and distribution comparisons
Preprocessing	KNIME (Partitioning, Missing Value, Normalizer, One to Many)	Stratified split, imputation, z-scoring, one-hot encoding
Modeling	KNIME Learner nodes, Python (Scikit-learn)	KNN, SVM, Random Forest, soft-voting ensemble
Evaluation	KNIME Scorer, Python	Macro-F1, Cohen's kappa, accuracy, and per-class recall/precision