Recomendação de quantificadores utilizando meta-aprendizado

Abstract

Metalearning based recommender systems have become an effective approach for the automatic selection of algorithms in machine learning tasks. This approach has been successfully employed in several studies to build recommenders across different domains. In the literature, there are works that investigate the recommendation of classifiers, hyperparameter configurations, and feature selection algorithms. Although a wide range of recommendation approaches has been proposed for classification problems, there is a lack of work in the context of quantification. Recently, quantification has attracted increasing interest in the machine learning community. Unlike classification, whose goal is to assign labels to unlabeled instances, quantification aims to estimate the class distribution in an unlabeled dataset. Several quantification methods have been proposed over the last years and, as in classification, there is no quantifier that consistently outperforms the others across all scenarios. Moreover, the empirical evaluation of quantifiers is particularly costly, as it requires generating multiple test sets with controlled prevalences. In this setting, determining the best quantifier for a given dataset is critical, yet labor-intensive when based on trial-and-error strategies. To address this gap, this work proposes the first meta-learning architecture for quantifier recommendation. The architecture models the relationship between dataset characteristics and quantifier performance to recommend a ranking of the most suitable quantifiers, without requiring trial-and-error experimentation. To evaluate the architecture, two quantifier recommendation strategies are developed. Experimental results demonstrate its effectiveness: under the best strategy, the system places the best-performing quantifier at the top of the recommended ranking in 91 out of 100 evaluated datasets. In addition, the ranking is directly explored in quantification tasks. When applying only the top-ranked quantifier, the method achieves performance comparable to the state-of-the-art DyS quantifier across the 100 datasets. On more complex datasets, performance remains stable and comparable to the best quantifier. Finally, combining the top quantifiers in the recommended ranking yields performance comparable to the state of the art over all datasets and superior performance when considering only the more complex datasets.