Original title: Learning Sparse Decision Trees via Transformer Variational Auto-Encoders
Giacomo Fidone, Alessio Cascione, Riccardo Guidotti · September 2026
Decision trees are popular machine learning tools because people can easily trace their logic. This makes them great for high-stakes choices. However, most training methods only focus on making accurate predictions. They ignore other important traits, like keeping the trees small and simple, which is called structural sparsityMaking a decision tree simpler by having fewer branches or nodes.. The researchers created a new method called TREVIS to fix this. It uses a Tree Transformer Variational Auto-EncoderA type of AI model that learns to compress and recreate data, used here to map trees. to turn whole decision trees into points in a smooth, continuous mathematical space. This allows the system to use standard calculus tools to search for better models smoothly. When tested, TREVIS found trees that matched the predictive power of existing near-optimal algorithms while making the trees simpler and sparser.
Imagine packing a suitcase where each item's position affects what you can fit next. Normally, you make discrete choices, and getting everything neat without wasting space is tricky. Here, the system learns to arrange items in a smooth space where small adjustments automatically balance neatness and capacity.
Decision trees guide critical choices in fields like healthcare, but they often become too complex to understand. This method keeps them accurate while making them simpler, so decisions are clearer and more trustworthy.
The abstract does not specify how TREVIS performs on very large or noisy datasets, nor does it explore objectives beyond sparsity and accuracy.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Learning Sparse Decision Trees via Transformer Variational Auto-Encoders — Giacomo Fidone, Alessio Cascione, Riccardo Guidotti
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.