Kernel-based methods are useful for various machine learning tasks. A kernel is a symmetrical positive definite function constructed on the graph of the data. Spectral analysis of the kernel can lead to an efficient representation. Such representation enables to reduce the size (dimension) of objects in complex datasets while preserving the coherence of the original data, such that clustering, classification, manifold learning and many other data analysis tasks can be applied in the reduced space. In this talk, I will describe two frameworks for fusing kernels from multiple views to extract a reliable, consistent representation from high-dimensional data sets. A new method for setting the kernel’s bandwidth will be presented, as well as applications for manifold learning, clustering, classification and detection of seismic events.