Ensemble Churn Classification Models
Budget: $30 – $250 USD
I have a telecom-style churn dataset that needs a full machine-learning pipeline, from preprocessing through model comparison. The raw file contains some missing values, several categorical columns, and features with different scales.
Here is what I need done:
• Data preparation – impute the gaps, one-hot (or target) encode the categoricals, and standardize or normalize the numeric features so every algorithm receives a clean, comparable matrix.
• Class-imbalance strategy – any sensible approach works (SMOTE, class weighting, etc.) as long as minority churn cases are properly represented during training and validation.
• Model building – train three ensemble learners: Random Forest, Gradient Boosting, and AdaBoost. For each, run a systematic hyper-parameter search (grid, randomized, or Bayesian) with cross-validation to squeeze out the best performance.
• Evaluation – report Accuracy, Precision, Recall, and ROC-AUC on a separate hold-out test set. Please include the confusion matrix and ROC curves to make the performance story visually clear.
• Comparison & recommendation – summarise which model you would deploy in production and why, backed by the metrics above.
Deliverables:
1. Clean, well-commented Python notebook or .py script (scikit-learn, imbalanced-learn, pandas, NumPy, matplotlib/seaborn).
2. A brief PDF/Markdown summary highlighting preprocessing choices, tuned parameters, and metric tables.
3. Saved model objects or joblib files so I can reproduce the results quickly.
I’ll consider the job complete when I can rerun the notebook end-to-end, reproduce the metrics on my machine, and clearly see which ensemble wins the churn challenge.
Here is what I need done:
• Data preparation – impute the gaps, one-hot (or target) encode the categoricals, and standardize or normalize the numeric features so every algorithm receives a clean, comparable matrix.
• Class-imbalance strategy – any sensible approach works (SMOTE, class weighting, etc.) as long as minority churn cases are properly represented during training and validation.
• Model building – train three ensemble learners: Random Forest, Gradient Boosting, and AdaBoost. For each, run a systematic hyper-parameter search (grid, randomized, or Bayesian) with cross-validation to squeeze out the best performance.
• Evaluation – report Accuracy, Precision, Recall, and ROC-AUC on a separate hold-out test set. Please include the confusion matrix and ROC curves to make the performance story visually clear.
• Comparison & recommendation – summarise which model you would deploy in production and why, backed by the metrics above.
Deliverables:
1. Clean, well-commented Python notebook or .py script (scikit-learn, imbalanced-learn, pandas, NumPy, matplotlib/seaborn).
2. A brief PDF/Markdown summary highlighting preprocessing choices, tuned parameters, and metric tables.
3. Saved model objects or joblib files so I can reproduce the results quickly.
I’ll consider the job complete when I can rerun the notebook end-to-end, reproduce the metrics on my machine, and clearly see which ensemble wins the churn challenge.