Big Mart Sales Forecasting

This project aims to predict sales of products at different Big Mart outlets. It utilizes a comprehensive machine learning pipeline, including data preprocessing, exploratory data analysis, model training, hyperparameter tuning, and model evaluation.

1. Data and Libraries

The project utilizes the following libraries:

• numpy: For numerical operations.
• pandas: For data manipulation and analysis.
• os: For operating system related functionalities.
• datetime: For date and time operations.
• matplotlib & seaborn: For data visualization.
• scikit-learn: For machine learning models and evaluation metrics.
• catboost, lightgbm, xgboost: For gradient boosting models.
• pickle: For saving and loading models.

The input data is assumed to be a CSV file named Train.csv in the /content/ directory of a Google Colab environment.

2. Project Workflow

The project follows these steps:

1. Exploratory Data Analysis (EDA):

   • Examines unique values in each column.
   • Checks for duplicates.
   • Identifies missing values.
   • Gathers basic data information.

2. Data Cleaning:

   • Handles missing values in 'Item_Weight' using mean imputation based on 'Item_Identifier'.
   • Addresses missing values in 'Outlet_Size' using mode imputation based on 'Outlet_Type'.
   • Replaces zero values in 'Item_Visibility' with the column mean.
   • Corrects inconsistencies in 'Item_Fat_Content' categories.
   • Creates a new feature 'New_Item_Type' from 'Item_Identifier'.
   • Modifies 'Item_Fat_Content' for 'Non-Consumables'.
   • Creates 'Outlet_Years' feature representing the outlet's age.

3. Data Visualization:

   • Generates bar plots for various categorical features (e.g., 'Item_Fat_Content', 'Item_Type', 'Outlet_Establishment_Year').
   • Visualizes the distribution of 'Item_Weight' using a distplot.
   • Displays a correlation heatmap for numerical features.

4. Data Preprocessing:

   • Label encodes categorical columns.
   • Performs one-hot encoding on specified columns.
   • Splits the data into training and testing sets.

5. Model Training and Evaluation:

   • Trains various regression models (Linear Regression, Ridge, Lasso, Decision Tree, Random Forest, Extra Trees, LGBM, XGBoost, and CatBoost).
   • Evaluates models using cross-validation and R2 score.
   • Visualizes feature importance for tree-based models.

6. Hyperparameter Tuning:

   • Performs RandomizedSearchCV for Random Forest, LGBM, XGBoost, and CatBoost.
   • Prints best parameters, best scores, and R2 scores.
   • Displays distribution plots of residuals.

7. Model Selection and Saving:

   • Selects the best-performing model (CatBoost Regressor).
   • Saves the trained model using pickle for later use.

8. Model Testing:

   • Loads the saved model and evaluates its performance on the full dataset.

3. Observations and Insights

• Several features show a strong correlation with sales.
• Tree-based models generally perform better than linear models.
• Hyperparameter tuning improves model accuracy.
• Missing values were imputed using reasonable strategies.
• Some inconsistencies in data required cleaning before modelling.
• Categorical feature engineering added significant value.

4. Conclusion

The project successfully created a robust sales forecasting model using various regression techniques, with CatBoost regressor delivering the highest performance. The model is ready for deployment.