Data fusion and matching by maximizing statistical dependencies

Multi-view learning is a task of learning from multiple data sources where each source represents a different view of the same phenomenon. Typical examples include multimodal information retrieval and classification of genes by combining heterogeneous genomic data. Multi-view learning methods can be motivated by two interrelated lines of thoughts: if single view is not sufficient for the learning task, other views can complement the information. Secondly, learning by searching for an agreement between views may generalize better than learning from a single view. In this thesis, novel methods for unsupervised multi-view learning are proposed. Multi-view learning methods, in general, work by searching for an agreement between views. However, defining an agreement is not straightforward in an unsupervised learning task. In this thesis, statistical dependency is used to define an agreement between the views. Assuming that the shared information between the views is more interesting, statistical dependency is used to find the shared information. Based on this principle, a fast linear preprocessing method that performs data fusion during exploratory data analysis is introduced. Also, a novel evaluation approach based on the dependency between views to compare vector representations for bilingual corpora is introduced. Multi-view learning methods in general assume co-occurred samples for the