Let’s talk!

Machine Learning and Statistical Analysis

title: Machine Learning and Statistical Analysis description: London Bike Sharing demand prediction in R date: 2024-01-23

Machine Learning and Statistical Analysis

London Bike Sharing demand prediction in R

London bike sharing

Introduction

The London Bike Sharing Dataset provides a detailed view of bike rental demand across hourly time intervals, enriched with weather and calendar information. In this project, I use R to explore the dataset, engineer time-based features, and train multiple regression models to predict hourly bike rental demand (cnt). The workflow includes data cleaning, exploratory analysis, outlier handling, model training, and performance comparison.

Project Objectives

  1. Data preprocessing: clean data, handle missing values, and align variable types.
  2. Exploratory data analysis (EDA): identify trends, seasonality, and relationships with weather/time features.
  3. Feature engineering: extract year, month, day, and hour from the timestamp and prepare categorical features.
  4. Model development: train and compare multiple regression models in R.
  5. Evaluation: assess models using RMSE, MSE, and R-squared (with and without outliers).
  6. Insights: highlight key drivers of bike demand and practical takeaways.

Data Source

Core features used:

  • timestamp: observation time (used for feature extraction)
  • cnt: total hourly rentals (target)
  • t1, t2: temperature and “feels like” temperature
  • hum: humidity
  • wind_speed: wind speed
  • weather_code: weather condition category
  • is_holiday, is_weekend: calendar flags
  • season: meteorological season category

Tools and Technologies

  • R / RStudio
  • Key packages:
    • dplyr, tibble, lubridate (wrangling + time features)
    • ggplot2, scales, gridExtra (visualization)
    • caret (splits, preprocessing)
    • rpart, randomForest, xgboost (modeling)
    • skimr, PerformanceAnalytics (profiling & correlation)

Methodology

1) Data Loading and Feature Engineering

After loading the dataset, I extracted time features (year, month, day, hour) from timestamp, then removed the original timestamp from modeling features.

Technical report snippet

2) Data Quality Checks

I inspected missing values, verified variable types, and reviewed summary statistics before modeling.

Missing values and dataset checks

3) Exploratory Data Analysis

EDA focused on:

  • distribution of bike rentals (cnt) and outliers
  • relationships with t1, hum, and wind_speed
  • seasonality by month and hour (rush-hour patterns)
  • categorical impacts (holiday/weekend/weather/season)
EDA visuals and trends

4) Feature Selection Notes

Two important adjustments were made for more reliable modeling:

  • Dropped t2 after confirming it was highly correlated with t1 but less reliable for consistency.
  • Excluded year 2017 due to significantly fewer observations than 2015–2016 (avoids imbalance and improves generalization).
Dropping t2 and filtering year

Modeling and Evaluation

Train/Test Split + Scaling

The dataset was split into training and testing subsets (80/20). Numeric features were standardized where required, then recombined with categorical variables.

Models Compared

  1. Linear Regression (baseline)
  2. Decision Tree (rpart)
  3. XGBoost (xgboost)
  4. Random Forest (randomForest)

Metrics

  • MSE (Mean Squared Error)
  • RMSE (Root Mean Squared Error)
  • R-squared (explained variance)

Model performance was compared with and without outliers to evaluate sensitivity to extreme demand spikes.

Model comparison table Model comparison without outliers

Key Results

  • XGBoost delivered the strongest overall performance (lowest error, highest R²).
  • Random Forest was consistently competitive and robust.
  • Removing outliers improved all models, especially the ensemble methods.
XGBoost feature importance

Insights

From both EDA and model behavior, the strongest demand drivers were:

  • Time of day (commute peaks in the morning and late afternoon)
  • Temperature
  • Humidity
  • Season / weather condition

These signals are useful for operational planning—bike redistribution, staffing, and expected demand forecasting.

Conclusion

This project demonstrates an end-to-end machine learning workflow in R, applied to a real-world urban mobility dataset. By combining statistical thinking, careful preprocessing, and model benchmarking, I built predictive models capable of estimating hourly bike demand in London. Ensemble methods—especially XGBoost—performed best, and outlier handling proved essential for improving accuracy.

Source Code and Resources