ICRAT 2014. Tutorial on Stationarity and Metrics. Massimiliano Zanin and Samuel Cristobal
The prestigious 6th International Conference on Research in Air Transportation included a tutorial by Innaxis Researchers S.Cristobal and M. Zanin, based on previous activities done within Resilience2050 project. The following is the abstract of their tutorial session, titled “Providing insight on how to apply data science in aviation: stationarity and metrics”
The aviation sector gathers and stores a large amount of unstructured, heterogeneous data from different sources and from diverse natures: safety data and reports, flight plans, navigation data, airport data, radar tracks, etc. From airlines to ANSPs to airports, the challenge to collect information through different data sensors is growing exponentially. Nevertheless, the manner in which different stakeholders make use of this data has not evolved rapidly enough, leaving a large gap for improvement. In this talk, we will review two topics of relevance for the correct application of data analysis to air transport. The first of which being “stationarity”, i.e. the importance of analysing data sets with coherent characteristics in time and space. When the stationarity of the system under study cannot be guaranteed, the results obtained can be plagued with errors and inconsistencies: for instances, causality relationships may spuriously appear, not due to a real dynamical mechanism, but just as a consequence of changes in the system. This is especially relevant when trying to forecast the future behaviour of the system by means of historical data, as relationships between the past and the future are an essential ingredient. The second topic will cover metrics, with focus on representativeness and significance. We study how different metrics can be misleading when interpreting the results of an study and how providing a simple answer is not always possible. Practical examples from ongoing research projects will be provided, namely: passenger-centered metrics, delay propagation metrics and resilience metrics. Representativeness and significance of the metrics will also be discussed in the tutorial.