IBM's explainability toolkit: a Swiss Army knife with 18 blades

What it does AIX360 is IBM Research’s open-source Python toolkit for explaining machine learning models and datasets. It wraps algorithms spanning local and global, post-hoc and direct, tabular and time-series explanations—plus a couple of proxy metrics (faithfulness, monotonicity) to check whether those explanations are lying to you.

The interesting bit The project acknowledges a genuine problem: no single explainer works everywhere. Rather than pretending otherwise, it ships a taxonomy tree and guidance material to help you navigate the maze. That’s unusually honest for a corporate research toolkit.

Key highlights

• Covers data explanations (ProtoDash, DIP-VAE), local post-hoc (LIME, SHAP, CEM variants), global direct (Boolean rules, GLRM, Ripper), global post-hoc (ProfWeight), time-series adaptations, and even a 2024 certification method (Ecertify)
• Modular pip install via algorithm-specific keywords like [rbm,dipvae,tsice]
• Supports tabular, text, image, and time-series data per the README
• Docker image and Jupyter examples included
• IBM Research-backed, with Slack community and explicit calls for contributions

Caveats

• Dependency fragmentation is real: some algorithms lock to Python 3.6, others to 3.10, and you cannot mix them in one environment (e.g., contrastive + rbm is a no-go)
• The README warns the library is “still in development”
• Default pip install only gets you base dependencies; most useful algorithms require explicit extras

Verdict Worth a look if you’re doing XAI research or need to compare multiple explanation methods without rewriting glue code. Skip it if you want a single, batteries-included install—or if your production stack can’t tolerate Python version pinball.