Improving the Intrinsic Interpretability of Differentiable Decision Trees through Euclidean Feature-Concentration Regularisation
Differentiable decision trees (DDTs) support end-to-end optimisation and explicit decision paths, but dense oblique splits can obscure the role of individual input features.
Explainer · on the wire
The explainer for this story is still on the wire.
The claim, the mechanism, and the figures land here once generated.
Source · arXiv