Flight Delay Detection

Abstract

Identifying flight delays is difficult for airlines, reducing efficiency and customer happiness. Classical machine learning and deep learning are compared for flight delay prediction using historical flight and meteorological data. The main goals were to assess the models' efficacy in both paradigms and establish which model/algorithm was better at identifying delayed flights, especially in cases when one class predominated. The study used considerable feature extraction to include time of day, day of the week, and meteorological parameters like wind direction and weather conditions to model training inputs. Logistic Regression, Support Vector Machines, Random Forest, LightGBM, Convolutional Neural Networks, Recurrent Neural Networks, and Long Short-Term Memory networks are examined.

Random Forest, optimized LightGBM, Logistic Regression and other conventional machine learning algorithms outperformed deep learning models in F1-scores and minority class identification (delayed flights). Deep learning models like LSTM had 91% accuracy, but low recall and F1-scores for minority categories made them unable to identify delayed flights. Classical models excelled in all categories and were versatile. When feature extraction and class imbalance are managed well, conventional machine learning models beat advanced deep learning models in flight delay prediction. In future aviation analytics applications, hybrid and intelligent sampling can improve flight delay detection.