Feature attribution for realized model fit (rather than predictions). EDEF provides exact additive decomposition of predictive performance for linear, PyTorch, JAX, and tree-based models, including standard errors, grouped inference, and exact TreeIG decompositions for tree-based models.
