DTSemNet: A New Neural Representation for Hard Oblique Decision Trees
Researchers have introduced DTSemNet, a novel representation of hard oblique decision trees as neural networks that enables end-to-end training with standard gradient descent without approximations. Decision trees are widely used in safety-critical domains like medical diagnosis due to their interpretability, but training accurate oblique trees is challenging because of complex optimization landscapes and overfitting risks. Existing differentiable approaches rely on approximations such as probabilistic softening (soft DTs) or quantized gradients like the Straight-Through Estimator (STE). DTSemNet overcomes these limitations by providing a semantically equivalent and invertible mapping, eliminating the need for approximations in both classification and regression tasks. The work is detailed in arXiv preprint 2605.07837.
Key facts
- DTSemNet is a new representation of hard oblique decision trees as neural networks.
- It enables end-to-end training with standard gradient descent.
- It eliminates the need for approximations like soft DTs or STE.
- Decision trees are used in safety-critical domains such as medical diagnosis.
- Training accurate oblique DTs is challenging due to complex optimization and overfitting.
- The work is published on arXiv with ID 2605.07837.
- DTSemNet is semantically equivalent and invertible.
- It applies to both classification and regression.
Entities
Institutions
- arXiv