Predicting Airline On-time Performance

Abstract

Growth of population and everyday needs made the world a busy place and transportation has now become one of the basic needs for every human being. Airlines being the easiest and fast transportation mechanism for long distance has increased its popularity over the years thus making a remarkable growth in aviation industry. Still this growth is not sufficient enough to cater air traffic congestion which causes flight delays. Therefore, the primary objective of this research is to predict and measure the flight delays so that the airlines can improve their ontime performance. The passengers have the benefit of adjusting their time schedules based on these predictions. The research focuses on predicting the departure delay of airlines while investigating the share of weather-related delays. The research problem is addressed as classification and regression tasks. Binary classification approach is used to classify the flights into delayed and non-delayed classes while regression is used to predict the delay time of a flight. The experiments are carried out using five years’ worth of flight records. Data sampling, encoding and scaling like preprocessing techniques used to prepare the data for learning. Logistic Regression, Linear Regression, Random Forest, Decision Trees and Naïve Bayes classification are used as statistical models. A Feed Forward and Convolutional neural networks are considered for the deep learning models. Each of these models were then evaluated using their respective performance matrices. Finally, the research came to a conclusion with the highest performing Random Forest model for the classification task. Feed forward neural network is identified as the suitable model for the regression task. Convolutional neural network seems to be the second best option for both classification and regression tasks.