The development of StackUmami includes multiple steps, including feature extraction, baseline models construction, the fusing feature representation, the final meta-based model development. Ten different feature encoding schemes from multiple perspectives were employed, which cover compositional information, composition-transition-distribution information, physicochemical properties and chemical structure information. These feature descriptors were then fed to five popular ML algorithms for constructing 50 baseline models to generate 50 PFs based on the feature representation learning strategy. Subsequently, the 50 PFs were used to train and optimize the final meta-based model using 10-fold cross-validation tests. Finally, the GA-SAR algorithm was used to identify and mine m out of 50 PFs that can help to improve the predictive ability of the final meta-based model.