View on GitHub

High performance, large-scale regression

by Alessandra Cabassi and Junyang Wang

Airline data

In the following, we analyse the airline dataset, publicly available for download from The data contains records of all commericial flights within the USA, from October 1987 to April 2008. It can be downloaded as 22 separate csv files, each containing the data for one year. When unzipped, the files take up 12 GB.

Each column in the csv files corresponds to one of the following covariates. Among others:Year comprised between 1987 and 2008, Month, DayOfMonth, DayOfWeek expressed as integers (for the days of the week, 1 is Monday), CRSDepTime and CRSArrTime the expected arrival and departure local times in the hhmm format,UniqueCarrier the unique carrier code, FlightNum the flight number, TailNum the plane tail number, CRSElapsedTime expected flight time in minutes, ArrDelay arrival delay, in minutes, DepDelay departure delay in minutes, Origin origin IATA airport code, Dest destination IATA airport code, Distance in miles.

Using this information, we want to see if it is possible to predict whether a flight will be delayed or not, making use of the information available before the departure. Therefore, in what follows, we binarise the ArrDelay column, setting each value to True if the ArrDelay is greater than zero, and False otherwise. Using this variable as our response, we perform logistic regression on the other covariates. The goal is to be able to do out-of-sample prediction and identify which variables influence delays the most.

We conducted Logistic Regression on the airplane dataset using both Spark and Tensorflow, summary of the analysis can be found:

Airplane data Logistic Regression with Apache Spark

Airplane data Logistic Regression with Tensorflow