Híper Heurísticas para la Selección de Métodos de Aprendizaje Profundo en la Clasificación de Textos Automatizada

Resumen

In this paper, an evolutionary model, in the scope of automated machine learning, that learns selection hyper-heuristics for text classification is presented. A hyper-heuristic is a set of if-then rules that evaluate a set of meta-features, summarizing the data distribution of a dataset, to select the most adequate deep learning method for such a dataset. It is expected that datasets with similar distributions can use the same classification model, generalizing the selection process. The model initially creates a population of hyper-heuristics at random and then evolves them using specific mutation and crossover operators. During the evolution, each hyper-heuristic is evaluated for its classification performance with a training group of datasets. At the end of the evolution, the best hyper-heuristic is chosen and evaluated for classification with an independent group of datasets. The results indicate that the best hyper-heuristic generalizes well the selection process, by choosing adequate classification methods for the datasets; and reaches a better performance than two state-of-the-art automated machine learning systems.