Kernel Learning for Structured Multi-View Data

It frequently happens in machine learning problems that the information explaining the subject of interest to be obtained from different sources or modalities and many well-studied algorithms are proposed for multi-view problems. The choice of the algorithms however, should adapt to the particular properties of each multi-view problem. Inspecting brain connectivity networks represented in brain images is a prevalent way to determine subjects affected by the neurological disorders. From a machine learning perspective, brain images and their corresponding class are labeled data instances which can be used for learning purposes. Medical images obtained from different imaging techniques are considered to be multi-view structured instances. Tensorial representation is the method we use in this study to preserve the natural structure of these images. Since the connectivity information corresponding to each brain image is embedded in tensorial structure, it is highly important that the implemented learning algorithm to preserve the multi-way structure. we use a structure preserving kernel representation of the tensor instances for all views of the data which eventually is exploited in a multiple kernel learning algorithm that is an extension of support vector machines for multi-view data.