New 'Twoblock' decision trees match random forests with full interpretability
A tree that explains itself as well as a black-box model predicts...
Sven Serneels has released a paper proposing twoblock clustering trees (tbtree), a new class of interpretable regression trees designed for multivariate responses. Unlike traditional decision trees that use axis-aligned splits and constant leaf values, tbtree employs dense or sparse twoblock dimension reduction both to determine splits and to fit local multivariate linear models at each leaf. This makes the tree deterministic and highly interpretable while still being computationally efficient.
Beyond the tree structure itself, Serneels introduces a novel method for estimating the twoblock dimension-reduced space by maximizing coskewness, a technique that helps identify non-normal clusters in data. Simulations show the tree recovers piecewise linear regimes effectively, and two real-world examples demonstrate that tbtree can match the predictive performance of black-box models like random forests—while keeping every split and leaf model open for inspection. This bridges the gap between accuracy and interpretability.
- Twoblock clustering trees use dimension reduction (dense/sparse) for both splits and leaf models, yielding interpretable local linear regimes.
- A new coskewness-based estimator detects non-normal clusters, enabling the tree to handle complex data distributions.
- Matches random forest accuracy on real-world datasets while maintaining full transparency of all model components.
Why It Matters
Interpretable AI that rivals black-box performance reduces trust barriers in regulated industries like finance and healthcare.