Similarity Forests Revisited: A Swiss Army Knife for Machine Learning
[ 1 ] Instytut Informatyki, Wydział Informatyki i Telekomunikacji, Politechnika Poznańska | [ P ] employee
2021
chapter in monograph / paper
english
- decision trees
- random forests
- similarity forests
EN Random Forests are one of the most reliable and robust general-purpose machine learning algorithms. They provide very competitive baselines for more complex algorithms. Recently, a new algorithm has been introduced into the family of decision tree learners – Similarity Forests, aiming at mitigating some of the well-known deficiencies of Random Forests. In this paper we extend the originally proposed Similarity Forests algorithm to one-class classification, multi-class classification, regression and metric learning tasks. We also introduce two new criteria for split evaluation in regression learning. The results of conducted experiments show that Similarity Forests can be a competitive alternative to Random Forests, in particular, when high quality data representation is difficult to obtain.
08.05.2021
42 - 53
20
140