In supervised learning, we often face incomplete or noisy data, which presents a significant challenge for predictive models, as noise can greatly reduce their performance. In this thesis, we explored an approach to modelling noisy data based on a specific neural network architecture, known as an autoencoder, adapted for classification tasks. We provided a detailed description of the construction and training process of this model. In the experimental evaluation, we tested the performance of the adapted autoencoder on the problem of predicting coronary heart disease. We added noise of different types and intensities to the data and compared the performance of this model with a standard multilayer neural network. The findings showed that, in most cases, the autoencoder proved to be a more successful model.
|
