An interesting area of machine learning is methods for multi-view data. Classical multi-view literature typically exploits multiple views to reduce noise in data via conditional independence. However, the multi-view relationships themselves are underutilized as factors for analyzing the data. In this work, we investigate the usefulness of these relationships in descriptive analytics and inference. In Part I, we cover a problem in radiation detection about inference of latent variables that are shared dependencies of multiple views. We show how the views can be aggregated by filtering their inferences collectively using domain knowledge about their relationships. In Part II, we address the problem of modeling multi-view structure directly and applying such structure to descriptive analytics and inference. We propose a method for radiation detection that learns linear multi-view correlations and detects anomalies when these correlations are disrupted. Next, we extend this idea to nonlinear multi-view correlations by introducing a clustering method whose correlations are cluster-wise linear. This method is applied to perform descriptive analytics on a medical problem. Furthermore, we extend this work to classification and demonstrate it on a load monitoring problem. Lastly, we propose an extension to regression.
Artur Dubrawski (Chair)
Simon Labov (Lawrence Livermore National Laboratory)