Machine learning, neural and statistical classification
Machine learning, neural and statistical classification
A Comparison of Ranking Methods for Classification Algorithm Selection
ECML '00 Proceedings of the 11th European Conference on Machine Learning
ICML '97 Proceedings of the Fourteenth International Conference on Machine Learning
| Hi-index | 0.01 |
Several methods have been proposed to generate rankings of supervised classification algorithms based on their previous performance on other datasets [8,4]. Like any other prediction method, ranking methods will sometimes err, for instance, they may not rank the best algorithm in the first position. Often the user is willing to try more than one algorithm to increase the possibility of identifying the best one. The information provided in the ranking methods mentioned is not quite adequate for this purpose. That is, they do not identify those algorithms in the ranking that have reasonable possibility of performing best. In this paper, we describe a method for that purpose. We compare our method to the strategy of executing all algorithms and to a very simple reduction method, consisting of running the top three algorithms. In all this work we take time as well as accuracy into account. As expected, our method performs better than the simple reduction method and shows a more stable behavior than running all algorithms.