Improved Gaussian process classification via feature space rotation

In the pattern recognition problem, numerous studies have been conducted to choose an appropriate feature space for improved classification results. Instead of directly applying any transformations to the input feature data, we aim to adjust the data distribution by appropriately rotating the coordinate system of the original feature data (in the high dimensional space), so that the data distribution along each feature dimension changes accordingly to support better modeling for classification. In particular, in this paper we present a novel feature space rotation strategy for Gaussian process classification, where rotation is applied to pairwise dimensions at a time and only those rotations that make the data more consistent among different dimensions are kept, as determined by an effective consistency measure we define. In this way, the data are adapted to better fit into the isotropic Gaussian process kernel. Extensive experimental results on various types of data sets have demonstrated the effectiveness of our approach.

► Novel feature space rotation for improved Gaussian Process classification.
► Adjust data distribution by appropriately rotating the feature coordinate system.
► The Consistency Index (CI) to guide feature space rotation effectively.
► The feature data is adapted to better fit the isotropic Gaussian process kernel.
► Extensive results demonstrated the effectiveness of our approach.