Introduction: Machine learning (ML) applications for studying asteroid resonant dynamics are a relatively new field of study. Results from several different approaches are currently available for asteroids interacting with the z2, z1, M1:2, and ν6 resonances. However, one challenge when using ML to the databases produced by these studies is that there is often a severe imbalance ratio between the number of asteroids in librating orbits and the rest of the asteroidal population. This imbalance ratio can be as high as 1:270, which can impact the performance of classical ML algorithms, that were not designed for such severe imbalances.Methods: Various techniques have been recently developed to address this problem, including cost-sensitive strategies, methods that oversample the minority class, undersample the majority one, or combinations of both. Here, we investigate the most effective approaches for improving the performance of ML algorithms for known resonant asteroidal databases.Results: Cost-sensitive methods either improved or had not affect the outcome of ML methods and should always be used, when possible. The methods that showed the best performance for the studied databases were SMOTE oversampling plus Tomek undersampling, SMOTE oversampling, and Random oversampling and undersampling.Discussion: Testing these methods first could save significant time and efforts for future studies with imbalanced asteroidal databases.