Possiamo ignorare la natura composizionale dei dati composizionali usando gli approcci di deep learning?

Care must be taken not to simply apply multivariate data analysis methods to compositional data. For example, one can show that correlations are biased to be negative, and almost all statistical methods result in biased estimates when applied to compositional data. One way out is to analyze data methods from compositional data analysis, i.e. by carrying out a log-ratio analysis. This contribution has its focus on settings where only the prediction and classification error is important rather than an interpretation of results. In this setting it is well-known that classification and prediction errors are smaller with a log-ratio approach using traditional machine learning methods. However, is this also true when training a neural network who may learn the inner relationships between parts of a whole also without representing the data in log-ratios? This contribution give an indication on this matter using one real data set from chemical measurements on beers.