Learning Shared and Discriminative Information from Multiview Data

With the advent of the Internet-of-Things (IoTs) and Big Data movements, the rate of gathering as well as accumulating information nowadays is growing dramatically. Huge amounts of data that abound with heterogeneous features representing distinct perspectives of the same underlying patterns arise in various scientific fields. For instance, LIDAR signals, radar signals, and camera videos can be seen as three different views of a particular self-driving car. In signal processing, machine learning, statistics, and data science, multiview learning (a.k.a., data fusion/integration) is a popular field with well-documented analytical tools and wide-range application domains. In this chapter, we will introduce two algorithms which seek to find the shared and discriminative information from multiview data. First, graph multiview canonical correlation analysis (GMCCA) models will be introduced to unravel such shared information, which aims at searching for the low-dimensional representations from multiview data, while incorporating some graph prior information of shared latent components of the multiview data. Capitalizing on kernel methods, a generalization of GMCCA to capture the latent components of nonlinear data data is further established. A theoretical analysis of the GMCCA is provided to estimate the generalization error bound. Second, we will introduce discriminative principal component analysis (dPCA) to learn the unique subspace of one dataset (a.k.a., target data) relative to the other dataset (a.k.a., background data). Under some mild assumptions, dPCA can be shown to be optimal in terms of finding the discriminative subspace of the target data with respect to the background data. Moreover, some representative applications are provided to validate the effectiveness of the aforementioned models in discovering shared or discriminative knowledge of multiview data.