Dacon - Flight Delay Probability Estimation Model

Desciption & Background:

Dacon is a famous AI competition hosting platform in Korea. Every month there is a monthly competition hosted by Dacon themselves. This April's topic is about estimating the probability of a flight being delayed or not. Having previous flight history information what kind of insight can we look into? The first thing is the imbalanceness of the class distributions in the data. Generally, Data of NOT DELAYED class are more that DELAYED class, leading to bias model.

For this project, I used an ensemble of Gradient Boosting Decision Trees where the output probabilities of all models are averaged to get the final predictions. There are two outputs, each represent the probability of being delayed and the other for not delayed.

The code is in python and XGB, CatBoost, and LightGBM model are used as base learner.

Keywords: Machine Learning, GBDT, Ensemble, Imbalanced