Predict Lost Flights Connections: An Interpretable Machine Learning Approach

12/14/2021

In airlines, flight schedule optimization and passenger satisfaction are problems that profoundly impact the airline industry revenue every year. Missed connections are often a consequence of unexpected disruptions and the lack of preventive mechanisms that affect airlines' regular operations and image. This thesis proposes a new approach for models to classify the success of passengers' connections through an airline hub, focusing on interpretability. This issue is key to airline profitability since decision-makers often want to have hard evidence before taking action. The models were trained on data from TAP Air Portugal's passenger activity from 2019 and the beginning of 2020, along with some data from airport movements. We analyzed the data and did some feature engineering, including encoding some features and generating new samples to re-balance the dataset. In total, we studied five models, two non-interpretable plus three interpretable models. The overall accuracy of the interpretable models was not as good as the results from the non-interpretable models. However, when looking for critical metrics for imbalanced data, as this is the case, and the performance on the minority class, i.e., missed connections, the interpretable models had a performance close to the one seen in the best non-interpretable model. These metrics included the Recall on the minority class and the macro-average Recall of the classification task as a whole. All models suggested that the most critical feature is the time scheduled for the connection and all of them gave none to marginal importance to features such as age or gender.

Authors in the community:

• 

  H

  Hugo Miguel Silva Lopes

  ist186637

Supervisors of this institution:

• 

  R

  Rodrigo Martins de Matos Ventura

  ist13950

• 

  C

  Cláudia Alexandra Magalhães Soares

  ist152210

mechanical-engineering - Mechanical engineering

eng - English

Instituto Superior Técnico