The paper addresses how machine learning models produce predictions through interpretability methods. It introduces two novel model-agnostic approaches – VarImp and SupClus – employing local regression with weighted distance accounting for variable importance. VarImp generates per-instance interpretations for complex relationships, while SupClus interprets clusters of similar instances for simpler datasets. Testing across multiple quantitative metrics demonstrates these methods achieve equal or superior interpretations, particularly for high-dimensional problems containing irrelevant features and nonlinear feature-target relationships.