This repository is the implementation of Tree in Tree: from Decision Trees to Decision Graphs.
Visualization examples of the fitted TnT graph structures:
- (Optional) install sknetwork to enable visualization:
pip install scikit-network# S indicates the number of split nodes, which is an estimate of model complexity.
| Model | Dataset | # S | Accuracy |
|---|---|---|---|
| TnT | MNIST | 600 | 90.87±0.31 |
| CART | MNIST | 1.1k | 88.59±0.14 |
| TnT | Connect-4 | 864 | 78.85±0.46 |
| CART | Connect-4 | 931 | 77.23±0.01 |
| TnT | Letter | 1.2k | 86.62±0.02 |
| CART | Letter | 1.3k | 86.26±0.15 |
| TnT | Optical recognition | 174 | 86.32±0.24 |
| CART | Optical recognition | 193 | 85.56±0.46 |
| TnT | Pendigits | 125 | 92.61±0.53 |
| CART | Pendigits | 166 | 91.74±0.13 |
| TnT | Protein | 69 | 57.26 |
| CART | Protein | 76 | 55.30 |
| TnT | SenseIT | 198 | 80.48±0.42 |
| CART | SenseIT | 345 | 79.40 |
| TnT | USPS | 31 | 88.76±1.36 |
| CART | USPS | 109 | 87.35±0.15 |
If you use this code for research, please consider citing our paper:
@misc{zhu2021tree,
title={Tree in Tree: from Decision Trees to Decision Graphs},
author={Bingzhao Zhu and Mahsa Shoaran},
year={2021},
eprint={2110.00392},
archivePrefix={arXiv},
primaryClass={cs.LG}
}