Background: With savvier management teams, airlines are becoming more stable, more productive, and more profitable. The problems plaguing the aviation industry, however, have not gone away and have become more complicated instead. Schedule recovery is the process of recovery from these issues (also known as operating disturbances). The recovery solver from Jeppesen is a software tool that produces a set of solutions to solve these operational disruptions.

Objectives: In this research work, we review the literature related to disruptions in airlines to understand the state of the art of applying machine learning and decrease the recovery time. The primary goal of this research work is to analyze the Jeppesenairline system and recovery solver extensively, which plays an important role and is used when disturbances occur. In the case of a loss, the recovery solver provides several solutions. The operator can either solve it manually, use a solution created by the recovery solver, or use a combination to solve a disturbance. The research also focuses on identifying various machine learning algorithms that can be used to answer two questions: "Will the operator use the solver" and "If the operator uses the solver, which solution will he prefer"

Methods: First, a literature review is performed to classify effective machine learning algorithms and then consider the findings of the discovery that an experiment is conducted to test the chosen machine learning algorithms. Due to unbalanced classes in the dataset, an experiment is performed to generate a synthetic dataset that is similar to the ground truth. Various steps that are done in the experimentation phase like data collection, preprocessing and training are described in detail. We also test the performance of various algorithms for machine learning.

Results: The results are presented in conjunction with the literature review and the experiments performed to answer research questions. The performance of the models is then measured using different performance metrics.

Conclusions: We finish the research work with an overall review of sections in the paper. It can be inferred that neural network models and the SVM model do not significantly improve predictive performance compared to the XGBoost model by evaluating the results obtained and considering the real-world scenario this study aims at.