A Proposed Meta-learning Framework for Algorithm Selection Utilising Regression-based Landmarkers
Daren Ler, Irena Koprinska, Sanjay Chawla · 2006
In this paper, we present a framework for metalearning that adopts the use of regression-based landmarkers. Each such landmarker exploits the correlations between the various patterns of performance for a given set of algorithms so as to construct a regression function that represents the pattern of performance of one algorithm from that set. The idea is that the independents utilised by these regression functions – i.e. landmarkers – correspond to the performance of a subset of the given algorithms. In this manner, we may control the number of algorithms being landmarked; the more that are landmarked, the fewer independents or evidence we have to make those approximations, and less accurate the landmarkers are. We investigate the ability of such landmarkers in combination with metalearners to learn how to predict the most accurate algorithm from a given set. While our results show that the accuracy of the meta-learning solutions increases as the quality of the metaattributes improves; i.e. when less algorithm performance measurements are landmarked and instead evaluated as independents, we find that in general, the results are still poor. However, we find that when a simple sorting mechanism is instead employed, the results are quite promising.