Flight_Delay_Prediction

✈️ Flight Delay Prediction

A machine learning project that predicts flight arrival delays using a two-stage pipeline — combining a classifier to detect whether a flight will be delayed, and a regressor to estimate the delay duration in minutes.

---

📌 Overview

Flight delays are a major disruption in the aviation industry, affecting millions of passengers each year. This project builds a two-stage prediction pipeline using real-world flight data and weather data collected from 15 major US airports across 2016–2017.

• Stage 1 — Classification: Predict whether a flight will be delayed (binary: delayed / not delayed).
• Stage 2 — Regression: For flights predicted as delayed, estimate the arrival delay in minutes.

Weather conditions at both origin and destination airports are incorporated as features, given weather’s significant impact on flight delays.

---

🗂️ Project Structure

Notebook	Description
Pre-Processing.ipynb	Merges raw flight and weather datasets for 2016 & 2017; handles nulls
Sampling.ipynb	Addresses class imbalance using SMOTE (oversampling) and NearMiss (undersampling)
Classification.ipynb	Trains and evaluates classification models; selects best classifier
Regression .ipynb	Trains and evaluates regression models; builds the final pipeline
Project.ipynb	End-to-end project notebook

---

🏗️ Data

• Flight Data: US domestic flights from 15 major airports (2016–2017)
• Weather Data: Hourly weather observations at origin and destination airports
• Airports covered: ATL, CLT, DEN, DFW, EWR, IAH, JFK, LAS, LAX, MCO, MIA, ORD, PHX, SEA, SFO

The datasets are merged on airport ID and time (departure/arrival) to align weather conditions with each flight record.

---

⚙️ Methodology

1. Data Preprocessing

• Merged flight and weather data at both origin and destination airports
• Dropped rows with null values
• Combined 2016 and 2017 datasets

2. Handling Class Imbalance

Since non-delayed flights significantly outnumber delayed flights, two sampling strategies were evaluated:

• NearMiss (undersampling)
• SMOTE (oversampling) — improved recall for the delayed class

3. Classification Models

| Model | Notes | |—|—| | Decision Tree | Baseline classifier | | XGBoost | Gradient-boosted ensemble | | Random Forest ✅ | Selected — highest recall on delayed class | | Logistic Regression | Linear baseline |

4. Regression Models

| Model | Notes | |—|—| | Gradient Boosting | Robust ensemble regressor | | XGBoost ✅ | Selected — best R² (0.95), lowest RMSE & MAE | | Random Forest | Ensemble regressor | | Decision Tree | Baseline regressor |

5. Pipeline

Input Flight + Weather Features
        │
        ▼
Random Forest Classifier
  ├── Not Delayed → output: 0 min delay
  └── Delayed ──▶ XGBoost Regressor → output: predicted delay (minutes)

---

📊 Results

Pipeline Performance

Metric	Score
RMSE	18.56
MAE	14.15
R² Score	0.94

Scores may vary depending on the train/test split ratio.

Best Standalone Regressor (XGBoost with SMOTE)

Metric	Score
R² Score	0.95
RMSE	16.42
MAE	11.45

---

🛠️ Tech Stack

---

🚀 Getting Started

1. Clone the repository

   git clone https://github.com/AMVamsi/Flight_Delay_Prediction.git
   cd Flight_Delay_Prediction
   

2. Install dependencies

   pip install pandas numpy scikit-learn xgboost imbalanced-learn matplotlib seaborn jupyter
   

3. Run the notebooks in order

   Pre-Processing.ipynb → Sampling.ipynb → Classification.ipynb → Regression .ipynb
   

---

✅ Conclusion

The two-stage pipeline successfully predicts flight arrival delays. Key findings:

• SMOTE oversampling significantly improved the recall for the delayed class in the classifier.
• Random Forest was selected as the classifier due to its highest recall score on delayed flights.
• XGBoost Regressor was selected for the pipeline due to its strong performance (R² = 0.95, RMSE = 16.42, MAE = 11.45).
• The final pipeline achieved an R² score of 0.94, demonstrating strong predictive capability.

---

📄 License

This project is licensed under the MIT License.