Learning the kernel matrix via predictive low-rank approximations

TL;DR

The paper introduces Mklaren, an efficient algorithm for approximating multiple kernel matrices that scales linearly with data points and kernels, enabling large-scale multiple kernel learning with high accuracy and interpretability.

Contribution

Mklaren is a novel geometrically-based algorithm that approximates multiple kernels without full matrix access, outperforming existing methods in accuracy and speed.

Findings

Outperforms contemporary kernel approximation methods.

Achieves highest explained variance with fewer iterations.

Reduces run time by two orders of magnitude for large datasets.

Abstract

Efficient and accurate low-rank approximations of multiple data sources are essential in the era of big data. The scaling of kernel-based learning algorithms to large datasets is limited by the O(n^2) computation and storage complexity of the full kernel matrix, which is required by most of the recent kernel learning algorithms. We present the Mklaren algorithm to approximate multiple kernel matrices learn a regression model, which is entirely based on geometrical concepts. The algorithm does not require access to full kernel matrices yet it accounts for the correlations between all kernels. It uses Incomplete Cholesky decomposition, where pivot selection is based on least-angle regression in the combined, low-dimensional feature space. The algorithm has linear complexity in the number of data points and kernels. When explicit feature space induced by the kernel can be constructed, a mapping from the dual to the primal Ridge regression weights is used for model interpretation. The Mklaren algorithm was tested on eight standard regression datasets. It outperforms contemporary kernel matrix approximation approaches when learning with multiple kernels. It identifies relevant kernels, achieving highest explained variance than other multiple kernel learning methods for the same number of iterations. Test accuracy, equivalent to the one using full kernel matrices, was achieved with at significantly lower approximation ranks. A difference in run times of two orders of magnitude was observed when either the number of samples or kernels exceeds 3000.