Applying different machine learning classifiers on very large dataset to demonstrate the challenges in choosing the right machine learning classifier for a given problem

Abstract Flight delays impose very heavy costs on airlines, customer as well as the economy. Besides other benefits, Machine Learning classifiers present an opportunity to analyze organizational data and make improvements based on hitherto undiscovered or not so obvious patterns. There is scope for increasing understanding of which Machine Learning classifier will deliver the best results for a given problem. This study used a large dataset on yearly flight statistics of the year 1997 from Bureau of Transportation Statistics (BTS) of the U.S. Department of Transportation with over 5 Million rows of data and compared results of seven classifiers, namely, Linear classifier, Regression classifier, Neural Net, Random trees, CHAID, Tree-AS (using CHAID algorithm and based on Regression Tree) and GLM (Generalized Linear Model) classifier to find significant variation in results as regards Correlation in predicted versus actual delay as well as key predictors the Departure delay in flight is seen to occur in 5 out of 6 cases indicating that while Machine Learning classifiers do agree on certain important parameters, differences between their findings are no less significant.