Skip to content

Latest commit

 

History

History
182 lines (127 loc) · 6.82 KB

README.md

File metadata and controls

182 lines (127 loc) · 6.82 KB
Raw

M5-H2-Regression-Competition

Project Overview

This project focuses on building a regression model to predict crop yield ('yield') using a dataset with various agricultural metrics. We employ extensive data analysis, feature engineering, and model tuning to minimize the model's error. The final model is stored as a serialized object for easy reuse in production.

Key Features

  • Custom Preprocessing Pipelines: Includes transformers for feature selection, feature engineering, and outlier handling.
  • Extensive Feature Engineering: New features are crafted based on domain knowledge to improve model performance.
  • Pipeline Integration: A single unified pipeline to streamline preprocessing, feature engineering, and model training.
  • Hyperparameter Tuning: Optimized hyperparameters for the RandomForestRegressor using advanced techniques optuna.

Repository Structure

M5-H2-Regression-Competition/
├── notebooks/
│   ├── EDA.ipynb                # Exploratory Data Analysis
│   ├── Model.ipynb              # Model training and evaluation
│   └── model_explain.ipynb      # Model explainability and interpretation
└── data/
│   ├── train.csv                # train.csv model trained datset
│   └── test.csv                 # test csv
├── README.md                    # Project documentation
├── LICENSE                      # Project license (MIT)
├── requirements.txt             # Required Python packages
├── model.pkl                    # Final model

Installation

  1. Clone the repository:

    git clone https://github.com/UznetDev/Wild-Blueberry-Prediction.git

  2. Navigate to the project directory:

    cd M5-H2-Regression-Competition

  3. Install the dependencies:

    pip install -r requirements.txt

Usage

Loading and Using the Model

To use the model stored in model.pkl, follow these steps:

import dill as pickle
import pandas as pd

# Load the trained model
with open('model.pkl', 'rb') as f:
    model = pickle.load(f)

# Prepare your test data (ensure it matches the training data format)
# test_data must have ['seeds', 'fruitmass', 'fruitset', 'AverageOfUpperTRange']
test_data = pd.read_csv('data/test.csv')

# Make predictions
predictions = model.predict(test_data)
print(predictions)

Model Details

The model pipeline integrates several custom transformers and a tuned RandomForestRegressor:

class ColumnSelector(BaseEstimator, TransformerMixin):
    ...
class FeatureEngineer(BaseEstimator, TransformerMixin):
    ...
class OutlierReplacer(BaseEstimator, TransformerMixin):
    ...
    
model = Pipeline([
    ('column_selector', ColumnSelector(columns=['seeds', 'fruitmass', 'fruitset', 'AverageOfUpperTRange'])),
    ('outlier_replacer', OutlierReplacer()),
    ('feature_engineer', FeatureEngineer()),
    ('model', RandomForestRegressor(...))
])

Dataset

The dataset is loaded from train.csv and includes features related to crop characteristics and environmental conditions. Detailed exploratory data analysis (EDA) is documented in notebook/EDA.ipynb.

Feature Engineering

The project includes custom feature engineering steps, such as creating ratios and interactions between features (e.g., FruitToSeedRatio, fruitset_seeds). These are implemented within the FeatureEngineer class.

Models Used

The primary model used is a RandomForestRegressor with custom hyperparameters. The pipeline approach allows easy modification and extension of the model, making it robust for handling diverse datasets.

Hyperparameter Tuning

Hyperparameters for the RandomForestRegressor were optimized with settings such as:

  • max_depth=9
  • n_estimators=497
  • max_features=0.809
  • min_samples_split=10
  • min_samples_leaf=4
  • criterion='absolute_error'

These values were chosen to maximize model performance while preventing overfitting.

Evaluation

The model was evaluated using standard regression metrics, including RMSE, R^2 and MAE. Details on evaluation and insights are in notebook/Model.ipynb.

Model Explanation

To understand this model, we use two powerful model explainers: SHAP and Permutation Importance.

  1. Permutation Importance:

    • Purpose: Permutation importance measures the impact of each feature on the model’s accuracy. It works by shuffling each feature and observing how much the model’s accuracy decreases. This technique helps identify which features are most crucial to the overall performance of the model.
    • Usage: We calculate the importance of each feature using permutation_importance from sklearn.inspection.
    • Plot: The permutation importance plot ranks features by their influence on model accuracy, making it easy to see which features are essential for the model’s performance.

  2. SHAP (SHapley Additive exPlanations):

    • Purpose: SHAP values explain individual predictions by showing the impact of each feature on the model’s output. This method highlights how each feature contributes to specific predictions.
    • Usage: We use shap.TreeExplainer to analyze our model, showing the effect each feature has on the model output.
    • Plot: The SHAP summary plot provides a bar chart, showing the average importance of each feature across all predictions, offering insights into which features are most influential.

You can understand model in notebook/model_explain.ipynb

Usage

The model is pre-trained and saved as model.pkl. Load and run it directly to make predictions on new data without retraining.

Results

The final model achieved strong results on the provided dataset, making it suitable for practical yield predictions in agricultural applications.

Contributing

Contributions are welcome! If you'd like to improve this project, please fork the repository and make a pull request.

  1. Fork the repository.
  2. Create a new branch for your feature or bug fix: 
    git checkout -b feature-name
  3. Commit your changes: 
    git commit -m "Add a new feature"
  4. Push to your branch: 
    git push origin feature-name
  5. Open a pull request.

License

This project is licensed under the MIT License - see the LICENSE file for details.

Contact's

If you have any questions or suggestions, please contact:

Thank you for your interest in this project. We hope it helps in your journey to understand and predict smoking habits using data science!