Abstract
As companies employ a larger number of models, the problem of algorithm (and parameter) selection is becoming increasingly important. Two approaches to obtain empirical knowledge that is useful for that purpose are empirical studies and metalearning. However, most empirical (meta)knowledge is obtained from a, relatively small set, of datasets. In this paper, we propose a method to obtain a large number of datasets which is based on a simple transformation of existing datasets, referred to as datasetoids. We test our approach on the problem of using metalearning to predict when to prune decision trees. The results show significant; improvement when using datasetoids. Additionally, we identify a number of potential anomalies in the generated datasetoids and propose methods to solve them.

Abstract
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. © Springer-Verlag Berlin Heidelberg 2001.

Abstract

Abstract

Abstract

Abstract