Explainable AI (XAI) · decision trees

When a model decides on a loan, a diagnosis or a scholarship, the why matters as much as the what. Explainable AI (XAI) seeks models whose decisions can be audited. Decision trees are the ideal case: the prediction is its own explanation — a path of questions from the root to a leaf.

My postdoc work goes deep into decision trees and interpretable learning: not just how accurate a model is, but how readable, and how to diagnose its failures.

Try it

The demo trains a real C4.5 tree on synthetic credit-approval data. You can:

• Read the model: move the instance (what-if) and watch the path that justifies the decision.
• Choose the split criterion: twoing, information gain (ID3) or gain ratio (C4.5), and see how the cuts change.
• See the real failures: as depth grows, bloat appears (the tree explodes in nodes without improving on test) — pruning fixes it; and under class imbalance, accuracy is misleading while AUC reveals the real performance.

An accurate model that never detects the minority (the defaults) is useless: that's why interpretability and the right metrics aren't a luxury, they are the work.