LR_Delivery_Time_Estimation (Porter Food Delivery)

This repository contains a complete data science pipeline for predicting food delivery times using Linear Regression. The project explores operational, temporal, and order-based features to build an accurate, interpretable model aimed at enhancing customer satisfaction and optimizing fleet management for Porter.

---

Project Overview

This notebook-based project is built on real-world food delivery data. It follows a structured machine learning workflow that includes:

• Data loading and cleaning
• Feature engineering (e.g., time features, one-hot encoding)
• Exploratory Data Analysis (EDA)
• Outlier handling and normalization
• Linear regression modeling and performance evaluation
• Business-focused insights and recommendations

The objective is to predict time_taken (in minutes) for food delivery based on various features like order size, distance, number of dashers, and time of day.

---

Key Features

1. End-to-End Workflow

A full ML pipeline is implemented using Python and libraries such as pandas, matplotlib, seaborn, and sklearn.

2. Feature Engineering

Created features like:

• hour, day_of_week, isWeekend
• Dummy variables for store_primary_category and order_protocol
• Target variable: time_taken from order timestamps

3. Exploratory Data Analysis (EDA)

• Distribution plots, boxplots, scatter plots
• Correlation heatmap
• Insights on peak hours, order complexity, and delivery patterns

4. Outlier Detection & Treatment

• Used IQR-based clipping to retain most data while handling extreme values

5. Predictive Modeling

• Trained a Linear Regression model with R² ≈ 0.86
• Performed Recursive Feature Elimination (RFE) to identify top predictors
• Conducted residual analysis to validate model assumptions

6. Business Recommendations

• Emphasize order size in ETA estimates
• Dynamically allocate delivery partners during peak hours
• Use complexity-based pricing for better revenue planning
• Improve restaurant operations for large orders

---

Visualizations Included

• Histograms & KDE plots for numeric features
• Boxplots for outlier detection
• Scatter plots for feature-target relationships
• Correlation matrix heatmap
• Residual diagnostics (Q-Q plots, residual vs. predicted)
• Feature importance bar charts

---

Assumptions

• Distance and price features are on natural scales (e.g., kilometers, INR).
• Timestamp conversions are in the dataset's original timezone.
• No rows were deleted—extreme values were clipped, not removed.

---

Files in This Repository

• LR_Delivery_Time_Estimation_Starter_Manish_Atwal.ipynb – Jupyter Notebook with full implementation
• LR_Delivery_Time_Estimation_Starter_Manish_Atwal.pdf – Final report with executive summary, EDA, and business insights

---

Reference

This project is inspired by real-world delivery optimization use cases and follows standard machine learning and EDA methodologies.

---

Author

Manish Atwal