In this thesis, we aim to raise the importance of deep clustering methods for multi-view and missing data. We combine two state-of-the-art methods, the Variational Deep Embedding (VaDE) initially designed for deep probabilistic clustering of single-view data and the Partial Variational Autoencoder (VAE) with the Pointnet Plus (PNP) structure, a method to predict missing data points. More precisely, we extend VaDE by an additional VAE and examine two fusion techniques for creating a shared distribution between two data views. To handle missingness in both views, we integrate the Partial VAE with PNP, which enables the definition of a partial clustering objective that depends on observed data samples only. As a result, we propose the Partial Multi-View Variational Deep Embedding (Partial MV-VaDE), a deep probabilistic clustering model targeting multi-view and missing data. We evaluate the model’s performance in extensive experiments with numerous multi-view data sets for which we generate different amounts of missingness. We observe the model’s changes in cluster probabilities in more detail and compare the clustering results to several baseline methods.
