TREVIS Accepted at ICDM 2026 🇨🇳

I am happy to share that our paper “Learning Sparse Decision Trees via Transformer Variational Auto-Encoders” has been accepted at The 26th IEEE International Conference on Data Mining — ICDM 2026!

This work was developed together with Giacomo Fidone (fist author) and Riccardo Guidotti.

The paper introduces TREVIS (Tree REpresentations from Variational Inference in latent Space), a generative approach to Decision Tree learning based on the exploration of the latent space of a Tree Transformer Variational Auto-Encoder (TTVAE).

Decision trees are widely used in interpretable machine learning because their decision processes can be directly inspected by humans. However, learning a good tree often requires balancing predictive performance with other desirable properties, such as structural sparsity.

TREVIS approaches this problem from a different perspective. Instead of directly searching over the discrete space of decision trees, we first learn a continuous latent representation of trees through the TTVAE. This makes it possible to explore and optimize the resulting latent space using a differentiable surrogate model.

In our experiments, we use TREVIS to jointly optimize predictive performance and sparsity. The results show that the method can discover decision trees with predictive performance comparable to existing near-optimal approaches while producing structurally sparser models.

The implementation of TREVIS is available on GitHub.

See you in Shenyang! 🇨🇳