2023_BA_Project

This repository is a team project for the 2023 Fall Business Analytics course in Seoul Nataional University of Science and Technology. The project aims to analyze the Seoul Public Bike(따릉이) project and develop strategies for deficit reduction. The analysis involves using machine learning models, specifically LightGBM and XGBoost, to predict public bicycle usage patterns and optimize system efficiency.

2023_BA_Project

Project Overview

The Seoul Public Bike project has faced financial challenges, with deficits increasing over the years. The goal of this project is to leverage business analytics to understand the usage patterns, optimize system efficiency, and propose strategies for deficit reduction.

Objectives

Financial Analysis: Conduct a comprehensive analysis of the Seoul Public Bike project's financial status, identifying trends, and understanding the factors contributing to deficits.
Usage Pattern Analysis: Explore usage patterns of public bicycles, considering factors such as borrowed hour, borrowed day, and environmental conditions.
Model Development: Implement machine learning models, including LightGBM and XGBoost, to predict bicycle usage and optimize station-specific trends.
Ensemble Model: Combine the strengths of LightGBM and XGBoost through ensemble modeling to enhance prediction accuracy.
Deficit Reduction Strategies: Based on the analysis results, propose strategies to reduce the financial deficits associated with the public bicycle project.

Execution Environment

Python 3.11 or higher
Conda (for managing the virtual environment)

Model Execution

Model Versions

pandas==1.5.3
numpy==1.24.3
scikit-learn==1.3.0
lightgbm==4.1.0
xgboost==2.0.2

Create and Activate Conda Environment

Create Conda Environment:
```
conda create --name myenv python=3.11
```
Replace myenv with the desired environment name.
Activate Conda Environment:
```
conda activate myenv
```

Install Required Packages

conda install --file requirements.txt

This command installs the necessary packages specified in the requirements.txt file within the Conda environment.

Launch Jupyter Notebook

jupyter notebook

Now, open the Jupyter Notebook and navigate to the team7_ensemble_model.ipynb notebook to run the code under the "1. Load and Preprocess Data" section.

Ensure that you are using Python 3.11 or a higher version and have activated your Conda environment before installing the required packages.

1. Load and Preprocess Data

Load the Data:

import pandas as pd

# Load data
data = pd.read_csv('merged_data.csv', encoding='utf-8')

Select Relevant Features and Preprocess Data:

# Selected Features
selected_features = ['stn_id', 'borrowed_hour', 'borrowed_day', 'is_holiday', 'borrowed_num_nearby', '강수량(mm)', 'wind_chill', 'nearby_id', 'borrowed_date', 'borrowed_num']
data = data[selected_features]

# Label Encoding for Categorical Features
categorical_features = ['stn_id', 'nearby_id']
for feature in categorical_features:
    data[feature] = pd.factorize(data[feature])[0]

2. Training and Model Creation

2.1 Train LightGBM Model

import lightgbm as lgb

# LightGBM Parameters for Regression
lgb_params = {
    'objective': 'regression',
    'metric': 'rmse',
    'boosting_type': 'gbdt',
    'num_leaves': 80,
    'learning_rate': 0.05,
    'feature_fraction': 1.0,
    'device': 'gpu'
}

# Create training and test datasets
train_data_lgb = lgb.Dataset(X_train, label=y_train)
test_data_lgb = lgb.Dataset(X_test, label=y_test, reference=train_data_lgb)

# Train the LightGBM model
lgb_model = lgb.train(lgb_params, train_data_lgb, num_boost_round=10000, valid_sets=[test_data_lgb, train_data_lgb], callbacks=[
    lgb.early_stopping(stopping_rounds=3, verbose=100),
])

2.2 Train XGBoost Model

import xgboost as xgb

# XGBoost Parameters for Regression
xgb_params = {
    'objective': 'reg:squarederror',
    'eval_metric': 'rmse',
    'booster': 'gbtree',
    'learning_rate': 0.1,
    'max_depth': 13,
    'subsample': 0.8,
    'device': 'gpu'
}

# Create training and test datasets
train_data_xgb = xgb.DMatrix(X_train, label=y_train)
test_data_xgb = xgb.DMatrix(X_test, label=y_test)

# Train the XGBoost model
xgb_model = xgb.train(xgb_params, train_data_xgb, num_boost_round=10000, evals=[(test_data_xgb, 'eval')], early_stopping_rounds=3, verbose_eval=100)

3. Ensemble Prediction

# Combine predictions of both models for ensemble prediction
y_pred_ensemble = (lgb_model.predict(X_test, num_iteration=lgb_model.best_iteration) + xgb_model.predict(test_data_xgb)) / 2

# Evaluate the performance of the ensemble model
ensemble_rmse = mean_squared_error(y_test, y_pred_ensemble, squared=False)
ensemble_r2 = r2_score(y_test, y_pred_ensemble)

print(f'Ensemble Test RMSE: {ensemble_rmse}')
print(f'Ensemble Test R-squared: {ensemble_r2}')

4. Deficit Reduction Strategies

4.1 Analysis and Insights

Based on the ensemble model results, analyze patterns and insights obtained from the predictions.

4.2 Propose Strategies for Deficit Reduction

Considering the analysis, propose effective strategies to reduce the financial deficits associated with the Seoul Public Bike project.

5. Results Storage

Save the predictions in the new_data_with_predictions.csv file:

# Save Predictions
new_data.to_csv('new_data_with_predictions.csv', index=False, encoding='utf-8')

6. Model Performance

After training and evaluating the LightGBM and XGBoost models, here are the key performance metrics:

LightGBM Model

Metric	Training Value	Test Value
RMSE	1.9139	1.9659
R-squared	0.5621	0.5377

XGBoost Model

Metric	Training Value	Test Value
RMSE	1.7220	1.9135
R-squared	0.6455	0.5620

Ensemble Model

Metric	Training Value	Test Value
RMSE	1.5199	1.7171
R-squared	0.7128	0.6473

These metrics provide insights into how well the models are performing, and users can quickly assess the quality of predictions.

7. Additional Notes and Considerations

Include any additional details, configurations, or modifications needed for the code. Clarify that the 'device' parameter is optional and can be adjusted based on the user's environment.

7.1 Hyperparameter Tuning Details

For detailed information about the hyperparameter tuning process for XGBoost and LightGBM, including the configurations used and insights gained, please refer to the 23_BA_preprocessing repository.

The hyperparameter tuning results and analysis can be found in the Hyperparameter Tuning section of the 23_BA_preprocessing repository.

8. Dataset

8.1 Training Dataset

To replicate the analysis and run the code, you'll need the dataset file merged_data.csv. You can download it using the following link:

Download merged_data.csv

Place the downloaded file in the project's root directory before running the Jupyter Notebook.

8.2 Collecting Real-time Rental Data (Optional)

If you want to collect real-time Seoul Public Bike rental data for testing purposes, you can use the provided Jupyter Notebook:

따릉이 Real Data Collection Notebook

Follow the instructions in the notebook to collect real-time rental data. Note that this step is optional, and you can proceed with the analysis without real-time data collection.

jeonghyeonee/2023_BA_Project