Comparative Analysis and Full-Stack Deployment of Machine Learning Models for Flight Fare Prediction

Domestic flight ticket pricing in India is influenced by multiple dynamic factors including airline carrier, travel route, departure timing, flight duration, and number of intermediate stops. Retrieving accurate fare estimates remains challenging due to continuously changing market demand, seasonal fluctuations, and varying airline pricing strategies. Existing travel and booking platforms primarily display current ticket prices and provide limited support for predictive fare estimation and intelligent booking analysis. Traditional statistical approaches often fail to capture complex nonlinear relationships between journey attributes and ticket prices, reducing prediction reliability across diverse travel scenarios.

Feature ranking by both mutual information regression and Random Forest importance scores confirms that Duration_total_mins is the dominant predictor, contributing 50.5% of ensemble split importance. Six regression algorithms are benchmarked under identical 75/25 train-test conditions: Linear Regression, Decision Tree, k-Nearest Neighbours (k=5), Gradient Boosting, XGBoost, and Random Forest. XGBoost achieves the highest test-set R² of 0.8312 with a Root Mean Squared Error of 1,812.78 INR and a Mean Absolute Percentage Error of 13.36%. The Random Forest tuned via RandomizedSearchCV with three-fold cross-validation attains R² = 0.8282 and RMSE = 1,828.80 INR. The serialised model artefact is stored as a Pickle file for deployment in booking advisory applications.