Our invention is a generalizable approach for understanding machine learning models using the model inputs and outputs only. We leverage topological data analysis (TDA), which is a new field for understanding complex high dimensional data by simplifying it into human understandable shapes capturing the most salient structures in the data. Our analysis approach builds on the 'Mapper" technique in a novel way by designing a cover specifically for machine learning predictions. This cover scheme substantially differentiates our approach from the existing Mapper algorithm. This allows us to overcome the scalability limitations of Mapper, which require the cover to be low dimensional. While our approach is tailored specifically towards machine learning applications, it scales to many dimensions, i.e., prediction classes. Additionally, we developed an analysis approach called 'escape routes" to explain relationships between different regions in the topological space defined model predictions.