autoxgboost - Automatic tuning and fitting of xgboost.
-
Install the development version
devtools::install_github("ja-thomas/autoxgboost")
autoxgboost aims to find an optimal xgboost model automatically using the machine learning framework mlr and the bayesian optimization framework mlrMBO.
Work in progress!
| Name | Factors | Numerics | Classes | Train instances | Test instances |
|---|---|---|---|---|---|
| Dexter | 20 000 | 0 | 2 | 420 | 180 |
| GermanCredit | 13 | 7 | 2 | 700 | 300 |
| Dorothea | 100 000 | 0 | 2 | 805 | 345 |
| Yeast | 0 | 8 | 10 | 1 038 | 446 |
| Amazon | 10 000 | 0 | 49 | 1 050 | 450 |
| Secom | 0 | 591 | 2 | 1 096 | 471 |
| Semeion | 256 | 0 | 10 | 1 115 | 478 |
| Car | 6 | 0 | 4 | 1 209 | 519 |
| Madelon | 500 | 0 | 2 | 1 820 | 780 |
| KR-vs-KP | 37 | 0 | 2 | 2 237 | 959 |
| Abalone | 1 | 7 | 28 | 2 923 | 1 254 |
| Wine Quality | 0 | 11 | 11 | 3 425 | 1 469 |
| Waveform | 0 | 40 | 3 | 3 500 | 1 500 |
| Gisette | 5 000 | 0 | 2 | 4 900 | 2 100 |
| Convex | 0 | 784 | 2 | 8 000 | 50 000 |
| Rot. MNIST + BI | 0 | 784 | 10 | 12 000 | 50 000 |
Datasets used for the comparison benchmark of autoxgboost, Auto-WEKA and auto-sklearn.
| Dataset | baseline | autoxgboost | Auto-WEKA | auto-sklearn |
|---|---|---|---|---|
| Dexter | 52,78 | 12.22 | 7.22 | 5.56 |
| GermanCredit | 32.67 | 27.67 | 28.33 | 27.00 |
| Dorothea | 6.09 | 5.22 | 6.38 | 5.51 |
| Yeast | 68.99 | 38.88 | 40.45 | 40.67 |
| Amazon | 99.33 | 26.22 | 37.56 | 16.00 |
| Secom | 7.87 | 7.87 | 7.87 | 7.87 |
| Semeion | 92.45 | 8.38 | 5.03 | 5.24 |
| Car | 29,15 | 1.16 | 0.58 | 0.39 |
| Madelon | 50.26 | 16.54 | 21.15 | 12.44 |
| KR-vs-KP | 48.96 | 1.67 | 0.31 | 0.42 |
| Abalone | 84.04 | 73.75 | 73.02 | 73.50 |
| Wine Quality | 55.68 | 33.70 | 33.70 | 33.76 |
| Waveform | 68.80 | 15.40 | 14.40 | 14.93 |
| Gisette | 50.71 | 2.48 | 2.24 | 1.62 |
| Convex | 50.00 | 22.74 | 22.05 | 17.53 |
| Rot. MNIST + BI | 88.88 | 47.09 | 55.84 | 46.92 |
Benchmark results are median percent error across 100 000 bootstrap samples (out of 25 runs) simulating 4 parallel runs. Bold numbers indicate best performing algorithms.
The Automatic Gradient Boosting framework was presented at the ICML/IJCAI-ECAI 2018 AutoML Workshop (poster).
Please cite our ICML AutoML workshop paper on arxiv.
You can get citation info via citation("autoxgboost") or copy the following BibTex entry:
@inproceedings{autoxgboost,
title={Automatic Gradient Boosting},
author={Thomas, Janek and Coors, Stefan and Bischl, Bernd},
booktitle={International Workshop on Automatic Machine Learning at ICML},
year={2018}
}