Predictive Analytics for Bike Sharing Demand

Developed a predictive model on Bike Sharing dataset, achieving 90% accuracy in demand prediction with Random Forest.

View Code on GitHub

Predictive Analytics for Bike Sharing Demand
Python icon
Python
scikit-learn icon
scikit-learn
AutoML icon
AutoML
Pandas icon
Pandas
NumPy icon
NumPy

Bike Sharing Demand Prediction

This project utilizes H2O AutoML to predict daily bike rental counts, focusing on model accuracy and feature interpretability. Excluding ensemble models, the analysis uses SHAP and LIME to understand key drivers of bike rentals.

AutoML Process

  • H2O AutoML was employed to automatically train multiple models such as Random Forest, Linear Regression, and XGBoost.
  • The goal was to predict the total daily bike rentals based on features like temperature, humidity, and seasonality.

Feature Engineering and Data Preprocessing

  • The dataset was preprocessed with one-hot encoding applied to categorical variables, while numerical features were already normalized.
  • Features include season, year, month, holiday status, weather conditions, and normalized temperature, humidity, and windspeed.

Results and Metrics

  • The Random Forest model outperformed other models with the following metrics:
    • R²: 0.87
    • MAE: 471
    • RMSE: 716
  • Linear Regression achieved:
    • R²: 0.84
    • MAE: 583
    • RMSE: 722
  • Temperature, year, and humidity were identified as the most important factors influencing bike rental demand.

Model Interpretability

  • SHAP (SHapley Additive exPlanations) analysis revealed:
    • Year (yr_1), temperature (temp), and feeling temperature (atemp) had the highest impact on rental counts.
    • SHAP values were used to explain how individual predictions were made, providing insights into how feature values contributed to the overall rental predictions.

Data Visualization

  • Histograms and count plots were used to explore the distribution of bike rentals and feature values.
    • Temperature and working days were found to significantly influence the number of rentals, with moderate temperatures correlating with higher rentals.

Limitations

  • H2O AutoML's inability to compute SHAP values for ensemble models led to the exclusion of these models to maintain interpretability.
  • The focus remained on simpler models like Random Forest and Linear Regression for transparent analysis.

For more information, visit the Github Repository.

Reading Time: 2 min read