ml-ralph

# ML-Ralph Experiment Workflow for Customer Churn Prediction

## Orient
- **Objective**: Improve customer churn prediction for TechSolutions Inc. in the SaaS industry.
- **Data**: 5 years of customer interaction data, including usage patterns, support tickets, and billing history.
- **Tools**: TensorFlow, Scikit-learn, Weights & Biases.

## Research
- **Key Findings**: Customer churn is highly correlated with support ticket resolution time and feature usage frequency.
- **Relevant Papers**: "Predicting Customer Churn with Deep Learning" (2022), "Feature Importance in Churn Models" (2021).

## Hypothesis
- **Hypothesis**: A model combining LSTM for temporal patterns and XGBoost for feature importance will outperform traditional logistic regression.

## Execute
- **Data Preprocessing**: Normalized numerical features, encoded categorical variables, and handled missing values.
- **Model Training**: Trained LSTM and XGBoost models separately and then combined their predictions.

## Analyze
- **Performance Metrics**: Combined model achieved an AUC-ROC of 0.89, compared to 0.82 for logistic regression.
- **Feature Importance**: Support ticket resolution time and feature usage frequency were the top predictors.

## Validate
- **Cross-Validation**: Model performance was consistent across different validation sets.
- **Business Impact**: Potential to reduce churn by 15% if implemented.

## Decide
- **Recommendation**: Deploy the combined LSTM-XGBoost model for real-time churn prediction.
- **Next Steps**: A/B test the model's predictions against current churn rates.