Adapting the variational auto encoder for datasets with large amounts of missing values.

This research provides an adaptation to the Vari- ational Auto Encoder(VAE) aimed at handling missing values. This adaptation comes in the form of a modified loss function. different VAE configurations are trained and evaluated on real- world data. These configurations include, next to the normal and modified loss functions, different beta parameters and usage of beta annealing. Practical implementations of the VAE or parts of it include dimensionality reduction and classification using network pretraining or a data preprocessor. While the modified loss function has a big influence on the VAE during training and reconstructing tasks, it does not have any influence on the performance of classification when either used as a data preprocessor or as a pretrained network. Lower beta values are preferred for reconstruction tasks and classification using a pretrained network. Higher beta values are preferred when using the VAE as a data preprocessor for classification. Annealing has shown to have no usable influence in research.