Automatic selection and recommendation of machine learning algorithms

Abstract

The use of meta-learning for algorithm recommendation is a research field that has been extensively explored in recent years. Meta-learning systems for recommending algorithms can be divided into two groups: (a) systems that perform selection of algorithms or models based on meta-features; and (b) systems that search for the best possible classification algorithm in a given space of algorithms, an approach relatively less investigated than the previous one. In previous work by the proponent and his collaborators, three representative search-based meta-learning systems were developed for the automatic construction of algorithms. All of them used the evolutionary algorithm paradigm as a search method, but building different types of classification algorithms -- rule induction, decision tree induction, and Bayesian network induction. These three meta-learning systems, to the best of our knowledge, are the only meta-learning approaches for the automatic construction of algorithms in the literature. Thus, the idea of this research project is to advance in the area of automatic construction of machine learning algorithms, more specifically classification. Among the topics considered open in this area, it is possible to highlight: (I) development of approaches based on different search methods; (II) evolution of decision trees with multiple tests; (III) implementation of multi-objective approaches; (IV) automatic construction of similarity functions in the context of semi-supervised learning; (V) automatic evolution of deep neural networks; (VI) automatic construction of ensemble of classifiers; (VII) automatic construction of multi-target predictors; and (VIII) comparative study of meta-learning approaches for selection/configuration and construction of classification algorithms. (AU)