Redução de dimensionalidade em dados de clima com uso de Stacked Autoencoders.

Over the past few decades, databases have been growing exponentially. This increase is not only in terms of the quantity of data samples but also in the number of features describing the variables, making datasets increasingly massive. Due to this, it becomes necessary to simplify these data sets by reducing their dimensionality. Climatic data are examples of data that often have many samples and involved characteristics. These factors result in high dimensionality, which, in turn, affects computational cost and predictive capacity, hindering the search for patterns and knowledge discovery. Furthermore, dimensionality reduction will benefit the visualization and storage of large climatic data sets. Thus, this work aims to use an Artificial Neural Network (ANN) with an Autoencoder (AE) architecture, specifically a type called a Stacked Autoencoder (SAE), to compress input climatic data, creating a compact and lower-dimensional representation. After that, with the AE’s ability to reconstruct the input data from this latent representation, it was found that the dimensionality reduction is good enough to reconstruct the data. These reconstructed data were evaluated using the mean squared error (MSE), resulting in a value of 0.01605. With this result, it is possible to confirm that the ANN fulfills its role of reducing the climatic data set.