AIX360

Why we included this project

Explaining what a model actually decided is a messy job, and this IBM Research library is one of the few that tries to cover the whole thing in a single Python package: local and global post-hoc explainers like LIME and SHAP, contrastive explanations, ProtoDash for picking representative examples, and rule-based models that are interpretable by construction. It also goes beyond the usual suspects with time-series explainers and proxy metrics such as faithfulness and monotonicity, which tell you whether an explanation is trustworthy rather than just plausible. The bundled guidance and taxonomy tree help you weigh the data-versus-model, local-versus-global, and post-hoc-versus-direct trade-offs, which matters because no single explainability approach fits every case. It installs via pip, so wiring it into a tabular or text pipeline is straightforward.