Delays in air transport can be seen as the result of two independent contributions, respectively stemming from the local dynamics of each airport and from a global propagation process; yet, assessing the relative importance of these two aspects in the final behaviour of the system is a challenging task. We here propose the use of the score obtained in a classification task, performed over vectors representing the profiles of delays at each airport, as a way of assessing their identifiability. We show how Deep Learning models are able to recognise airports with high precision, thus suggesting that delays are defined more by the characteristics of each airport than by the global network effects. This identifiability is higher for large and highly connected airports, constant through years, but modulated by season and geographical location. We finally discuss some operational implications of this approach.