Grain boundaries (GBs) in natural gas hydrates (NGHs) are ubiquitous crystal imperfection that undermine the stability of NGHs. In this study, we report the diverse cage structures and mechanical properties of GBs in methane hydrates using molecular dynamic (MD) simulations with advanced machine learning (ML) techniques. Employing a Random Forest-based feature extraction method, we weight and rank the 84 degrees of freedom to eliminate cage types with zero weight. Notably, 11 cage types identified through principal component analysis (PCA) play critical roles in dictating the overall mechanical properties of GBs in methane hydrates. This work establishes an MD-ML framework to elucidate the relationships between GB composition-structure-mechanical properties in NGHs, which is of great importance in predicting the mechanical stability of sediment-hosted NGHs.