Meta-learning framework for prediction strategy evaluation

Abstract

The paper presents a framework which brings together the tools necessary to analyze new problems and make predictions related to the learning algorithms' performance and automate the analyst's work. We focus on minimizing the system dependence on user input while still providing the ability of a guided search for a suitable learning algorithm through performance metrics. Predictions are computed using different strategies for calculating the distance between datasets, selecting neighbors and combining existing results. The framework is available for free use on the internet. © 2011 Springer-Verlag Berlin Heidelberg.