Optimizing Sales Prediction for Rohlik with Machine Learning and Weighted MAE Evaluation

Libraries Used

Data Handling & Computation:

• pandas: Used for data manipulation and preprocessing.
• numpy: Provides numerical computing capabilities, especially for handling arrays and mathematical functions.

Visualization:

• matplotlib: Used for creating static, animated, and interactive visualizations.
• seaborn: Built on top of Matplotlib, provides enhanced data visualization features.

Machine Learning:

• scikit-learn: Includes tools for machine learning models, preprocessing, feature selection, and evaluation metrics.

Model Serialization:

• joblib: Utilized for saving and loading trained models efficiently.

APIs & Deployment:

• Flask: Used for building APIs to serve machine learning predictions.
• Docker: A containerization tool used for deploying the API in a scalable and reproducible environment.

---

Workflow Summary

Data Preparation

Handling Missing Values:

• Filled NaN values in holiday_name with "No Holiday" to ensure continuity in data.

Feature Engineering:

• Extracted date-related features such as year, month, day, weekday, and weekend indicators.
• Created lag and rolling features like lag_1 and rolling_mean_7 to incorporate past trends.
• Added holiday indicators to account for special dates impacting sales.

---

Model Training & Evaluation

Regression Models:

• Linear Regression: Evaluated using MAE, MSE, and RMSE.

Classification Models:

• Logistic Regression: Assessed based on accuracy and confusion matrix.

Key Metrics:

• MAE (Mean Absolute Error): Measures the average magnitude of errors in predictions.
• MSE/RMSE (Mean Squared Error & Root Mean Squared Error): Penalizes large errors more significantly.
• MAPE/SMAPE (Mean Absolute Percentage Error & Symmetric MAPE): Useful for percentage-based errors.
• R-squared: Indicates how well the model explains variability in the data.
• WMAE (Weighted Mean Absolute Error): Assigns more importance to specific observations, like holiday sales.

---

Optimization

Hyperparameter Tuning:

• Applied GridSearchCV and RandomizedSearchCV for optimal parameter selection.

Feature Selection:

• Techniques included Recursive Feature Elimination (RFE), SelectKBest, and correlation analysis.

Ensemble Methods:

• Utilized RandomForest and GradientBoosting for improving model performance.

Regularization:

• Applied L1/L2 regularization to prevent overfitting.

---

Deployment

Model Saving:

• Serialized models using joblib for easy reusability.

API Creation:

• Developed a Flask API to serve predictions efficiently.

Containerization:

• Packaged the application into a Docker container for deployment.

Cloud Deployment:

• Hosted the API on cloud platforms like Heroku to enable accessibility.

Visualization Techniques

• Actual vs. Predicted Plots: Bubble charts where size/color represents weights.
• Error Distribution: Visualized using bar plots and boxplots.
• Residual Analysis: Scatter plots showing residuals vs. actuals.
• Metric Comparison: Comparative bar plots for MAE vs. WMAE.

---

Key Takeaways

• Focus on WMAE: Prioritizes key observations such as holiday sales.
• Optimization Matters: Feature selection and hyperparameter tuning enhance model accuracy.
• Deployment Readiness: APIs and Docker ensure scalability and accessibility.