Non-linear distance based large scale data classifications

摘要

Linear subspace projections are an important technique to reduce the dimensionality of data for automatic classification. Especially for large-scale and on-line systems, e.g. gesture recognition applications, this is important to guarantee near real-time processing. The linear subspace projections, however, fail if the classes are not linearly separable. Kernel methods, in contrast, have been widely applied to linear classification algorithms to solve problems of non-linearly separable classes. This technique, however, increases the computational complexity by introducing the evaluation of a possibly non-linear function. Here, we extend a linear subspace projection that has been applied to large-scale systems using a kernel function. The method is evaluated on Fisher's Iris dataset and a recorded gesture dataset. The results indicate that the proposed method yields an increased accuracy at a subspace of lower dimension while achieving a similar runtime at a subspace of the same dimension. The proposed method is thus expected to work well with online systems.