Algorithm Selection via Meta-Learning and Active Meta-Learning

To find most suitable classifier is possible either through cross-validation, which suffers from computational cost or through expert advice which is not always feasible to have. Meta-Learning can be a better approach to automate this process, by generating Meta-Examples which is a combination of performance results of classification algorithms on input datasets and Meta-Features. With the increasing number of datasets and underlying complexity of algorithms, makes even the Meta-Learning process expensive. So, Active Meta-Learning can help by optimizing the generation of Meta-Examples along with maintaining the performance of classification algorithms. Proposed work here provides a ranking of classifiers using SRR and ARR ranking method and compares Meta-Learning with Active Meta-Learning. In this work, evaluation methodology based on ideal ranking is presented, which shows that proposed method leads to significantly better ranking with reduced Meta-Examples. The executed experiments discovered a considerable improvement in Meta-Learning performance that supports nonexperts users in the selection of classification algorithms.