Choosing the proper autoencoder for feature fusion based on data complexity and classifiers: Analysis, tips and guidelines

摘要

Classifying data patterns is one of the most recurrent applications in machine learning. The number of input features influences the predictive performance of many classification models. Most classifiers work with high-dimensional spaces. Therefore, there is a great interest in facing the task of reducing the input space. Manifold learning has been shown to perform better than classical dimensionality reduction approaches, such as Principal Component Analysis and Linear Discriminant Analysis. In this sense, Autoencoders (AEs) provide an automated way of performing feature fusion, finding the best manifold to reconstruct the data. There are several models and architectures of AEs. For this reason, in this study an exhaustive analysis of the predictive performance of different AEs models with a large number of datasets is proposed, aiming to provide a set of useful guidelines. These will allow users to choose the appropriate AE model for each case, depending on data traits and the classifier to be used. A thorough empirical analysis is conducted including four AE models, four classification paradigms and a group of datasets with a variety of traits. A convenient set of rules to follow is obtained as a result.