Multiple Outlooks Learning with Support Vector Machines

Multiple Outlooks Learning (MOL) has recently received considerable attentions in machine learning. While traditional classification models often assume patterns are living in a fixed-dimensional vector space, MOL focuses on the tasks involving multiple representations or outlooks (e. g., biometrics based on face, fingerprint and iris); samples belonging to different outlooks may have varying feature dimensionalities and distributions. Current MOL methods attempted to first map each outlook heuristically to a common space, where samples from all the outlooks are assumed to share the same dimensionality and distribution after mapping. Traditional off-the-shelf classifiers can then be applied in the common space. The performance of these approaches is however often limited due to the independence of mapping functions learning and classifier learning. Different from existing approaches, in this paper, we proposed a novel MOL framework capable of learning jointly the mapping functions and the classifier in the common latent space. In particular, we coupled our novel framework with Support Vector Machines (SVM) and proposed a new model called MOL-SVM. MOL-SVM only needs to solve a sequence of standard linear SVM problems and converges rather rapidly within only a few steps. A series of experiments on the 20 newsgroups dataset demonstrated that our proposed model can consistently outperform the other competitive approaches.