Meta-learning

In several domains, such as in Machine Learning, there is a variety of algorithms that can be considered as candidates to solve particular problems. One of the most difficulty tasks in these domains is to predict when one algorithm is better than another to solve a given problem. Traditional approaches to predicting the performance of algorithms often involve costly trial-and-error procedures. Other approaches require expert knowledge, which is not always straightforward to acquire.

In the previous context, meta-learning approaches have arisen as effective solutions, able to automatically predict algorithms performance for a given problem. Thus, such approaches could support non-expert users in the algorithm selection task. There are different interpretations for the term “meta-learning”. In our work, we use “meta-learning” meaning the automatic process of generating knowledge that relates the performance of machine learning algorithms to the characteristics of the problem (i.e., characteristics of its datasets).

So far, in the literature, meta-learning had been used only for selecting/ranking supervised learning algorithms. Motivated by this, we extend the use of meta-learning approaches for clustering algorithms. We developed our case study in the context of clustering algorithms applied to cancer gene expression data generated by microarray.

There are several works that can be developed from our proposal by, for example, implementing other meta-learners for different categories of datasets, and by using other meta-learning approaches that have not yet been used in the algorithm selection problem. Recently, in the context of supervised learning, we analyzed the relevance of the complexity measures as meta-attributes.

Main publications:

Luís P. F. Garcia, Ana Carolina Lorena, Marcilio C. P. de Souto, and Tin Kam Ho. Classifier recommendation using data complexity measures. In ICPR, pages 874–879. IEEE Computer Society, 2018. doi: 10.1109/ICPR.2018.8545110
Marcílio C. P. de Souto et al. Ranking and selecting clustering algorithms using a meta-learning approach. In IJCNN, pages 3729–3735, 2008. doi: 10.1109/IJCNN.2008.4634333
Ricardo Prudêncio, Marcilio C. P. de Souto, and Teresa B. Ludermir. Selecting machine learning algorithms using the ranking meta-learning approach. In Meta-Learning in Computational Intelligence, volume 358 of Studies in Computational Intelligence, pages 225–243. Springer, 2011. ISBN 978-3-642-20979-6. doi: 10.1007/978-3-642-20980-2\_7