flight-delay-predictor

https://drive.google.com/file/d/16rbiBsJlo9gm5-Mq28ctsXCEuzfLZA5E/view?usp=drive_link

Flight Delay Prediction — Full Project (Parts 1-8) -

Overview -

• This project analyzes U.S. domestic flight data and builds both regression and classification models to predict flight arrival delays.
• The work is structured into eight clear stages, following the assignment’s required workflow.

1. Dataset Overview -

• The dataset contains ~96K domestic U.S. flights. It includes:
• Scheduling information (YEAR, MONTH, DAY, SCHEDULED_DEPARTURE, etc.)
• Operational details (DISTANCE, AIRLINE, ORIGIN_AIRPORT, DESTINATION_AIRPORT)
• Delay-related causes (WEATHER_DELAY, NAS_DELAY, LATE_AIRCRAFT_DELAY, etc.)
• Target variable: ARRIVAL_DELAY (minutes)
• Goal: Build predictive models to estimate arrival delay and explore the key operational factors affecting punctuality.

1. Exploratory Data Analysis (EDA) -

• Main steps performed:
• Checked missing values → Only a few columns contained missing values, and all were handled explicitly.
• Identified relevant delay columns and analyzed their contribution (weather, NAS, late aircraft).
• Examined seasonality and time-of-day patterns (monthly delays, hourly delays).
• Compared delays between airlines.
• Visualized relationship between distance and delay.
• Key Findings:
• Certain months show heavier congestion.
• Evening flights have systematically higher delays (“snowball effect”).
• Airlines differ strongly in punctuality.
• Distance has almost no explanatory power.

1. Baseline Regression Model -

• Steps completed:
• Removed leakage features (e.g., DEPARTURE_DELAY, WHEELS_ON).
• Used only information available before takeoff.
• Trained a simple Linear Regression model.
• Evaluated the model using MAE, MSE, RMSE, and R².
• Results (Baseline):
• RMSE ≈ 9.23 minutes
• R² ≈ 0.88
• Train/test scores were close - no overfitting.

1. Feature Engineering -

• Performed multiple transformations to enhance model performance:
• 4.1 Encoding -
• One-Hot Encoding - AIRLINE
• Frequency Encoding - ORIGIN_AIRPORT, DESTINATION_AIRPORT
• 4.2 New Features -
• IS_WEEKEND — captures weekend travel differences
• 4.3 Clustering -
• Applied K-Means (k=4) on DISTANCE and SCHEDULED_TIME
• Added new feature: FLIGHT_CLUSTER
• 4.4 Dimensionality Reduction -
• PCA visualization to validate that clusters form meaningful groups
• 4.5 Scaling -
• Removed leakage / irrelevant fields
• Scaled the final 33 features using StandardScaler

1. Improved Regression Models -

• Trained three models on the engineered dataset:
• Linear Regression (Improved)
• Random Forest Regressor
• Gradient Boosting Regressor
• Results:
• Gradient Boosting achieved best performance
• RMSE ≈ 9.04
• R² ≈ 0.89
• This improves over the baseline because tree-based models capture non-linear relationships.

1. Winning Regression Model + Deployment -

• Selected Gradient Boosting Regressor as winner.

• Exported it using pickle:

• with open("winning_model.pkl", "wb") as f:

  • pickle.dump(best_model, f)

• Uploaded to a dedicated HuggingFace model repository.

1. Regression - Classification -

• 7.1 Creating Classes -
• Converted arrival delay into 3 classes using quantile binning:
• Class 0: lowest 33% delays
• Class 1: middle 33%
• Class 2: highest 33%
• Why?
• This ensures balanced classes and avoids distortions caused by skewed delay distributions.
• 7.2 Class Balance Check -
• Class distribution remained balanced (≈ 32–36% per class).
• Therefore:
• Accuracy is meaningful
• Also tracked macro-F1 to ensure fair performance across classes

1. Classification Models -

• 8.1 Precision vs Recall -
• We evaluated both but emphasized recall because misclassifying high-delay flights as low-delay is more costly than the opposite.
• 8.2 Models Trained
• Logistic Regression
• Random Forest Classifier
• Gradient Boosting Classifier
• All trained on the engineered features.
• 8.3 Evaluation
• For each model:
• Classification report (precision, recall, F1-score)
• Confusion matrix
• Analysis of error patterns
• Best model (macro F1): Logistic Regression
• Even though simple, it produced the most balanced performance across all classes.
• 8.4 Exporting the Winning Classifier
• with open("winning_classifier.pkl", "wb") as f:
  • pickle.dump(best_cls_model, f)
• Uploaded the classifier to the same HF repository as required.