Randomer Forests
share
Random forests (RF) is a popular general purpose classifier that has been shown to outperform many other classifiers on a variety of datasets. The widespread use of random forests can be attributed to several factors, some of which include its excellent empirical performance, scale and unit invariance, robustness to outliers, time and space complexity, and interpretability. While RF has many desirable qualities, one drawback is its sensitivity to rotations and other operations that "mix" variables. In this work, we establish a generalized forest building scheme, linear threshold forests. Random forests and many other currently existing decision forest algorithms can be viewed as special cases of this scheme. With this scheme in mind, we propose a few special cases which we call randomer forests (RerFs). RerFs are linear threshold forest that exhibit all of the nice properties of RF, in addition to approximate affine invariance. In simulated datasets designed for RF to do well, we demonstrate that RerF outperforms RF. We also demonstrate that one particular variant of RerF is approximately affine invariant. Lastly, in an evaluation on 121 benchmark datasets, we observe that RerF outperforms RF. We therefore putatively propose that RerF be considered a replacement for RF as the general purpose classifier of choice. Open source code is available at http://ttomita.github.io/RandomerForest/.
READ FULL TEXT