MetaLProjection: uma abordagem para recomendação de algoritmos de redução de dimensionalidade utilizando meta-aprendizagem

Abstract

Data visualization techniques have enormous potential for analysis, summarization, understanding in order to facilitate information extraction. Recent areas such as Visual Analytics and Data Science highlight its importance, especially for analyzing complex sets of data. In this sense, Multidimensional Projection techniques are particularly used to perform visual analysis of high-dimensional data sets, because they perform dimensionality reduction and, consequently, have better scalability regarding the number of attributes/dimensions. However, there is a wide variety of these projection techniques and defining which is most suitable for finding visual patterns of information in one, or several, data sets is not a trivial task. Although works in the literature test and compare different techniques on data sets with different characteristics, they do not do so in a systematic way to facilitate a choice for the user. In this context, this research uses meta-learning to classify and recommend multidimensional projections considering certain evaluation metrics in a knowledge base with more than 500 distinct data sets. To evaluate the approach, i) the relationship between the meta-attributes of all data sets was observed, ii) generation of a ranking containing the performance of the chosen projection techniques and, iii) performance of the recommendation of these techniques. Finally, the results obtained show that the developed approach contributes efficiently to the choice and recommendation of multidimensional projection techniques.