Abstract:The random Fourier features (RFFs) method is a powerful and popular technique in kernel approximation for scalability of kernel methods. The theoretical foundation of RFFs is based on the Bochner theorem (Bochner in Harmonic Analysis and the Theory of Probability, University of California Press, 1995) that relates symmetric, positive definite (PD) functions to probability measures. This condition naturally excludes asymmetric functions with a wide range applications in practice, e.g., directed graphs, conditional probability, and asymmetric kernels. Nevertheless, understanding asymmetric functions (kernels) and its scalability via RFFs is unclear both theoretically and empirically. In this paper, we introduce a complex measure with the real and imaginary parts corresponding to four finite positive measures, which expands the application scope of the Bochner theorem. By doing so, this framework allows for handling classical symmetric, PD kernels via one positive measure; symmetric, non-positive definite kernels via signed measures; and asymmetric kernels via complex measures, thereby unifying them into a general framework by RFFs, named AsK-RFFs. Such approximation scheme via complex measures enjoys theoretical guarantees in the perspective of uniform convergence. In algorithmic implementation, to speed up the kernel approximation process, which is expensive due to the calculation of total masses, we propose a subset-based fast estimation method. This method focuses on optimizing total masses within a sub-training set, effectively transforming the numerical integration for total masses into quadratic programming in three-dimension with low time complexity. AsK-RFFs provides two explicit feature mappings to approximate an asymmetric kernel. These mappings can be utilized in classifiers within the framework of AsK-LS (He et al. in IEEE Trans Pattern Anal Machine Intell 45(8):10044–10054), fulfilling the purpose of using asymmetric kernels in machine learning applications. Our AsK-RFFs method is empirically validated on several typical large-scale datasets and achieves promising kernel approximation performance, which demonstrates the effectiveness of AsK-RFFs.

Orthogonal Random Features: Explicit Forms and Sharp Inequalities

Towards Unbiased Random Features with Lower Variance for Stationary Indefinite Kernels

A Double-Variational Bayesian Framework in Random Fourier Features for Indefinite Kernels

Random Fourier Feature Kernel Recursive Least Squares.

Random fourier features for asymmetric kernels

Stein Random Feature Regression

All Random Features Representations are Equivalent

Random Fourier Features for Operator-Valued Kernels

New random projections for isotropic kernels using stable spectral distributions

Adaptive and Oblivious Randomized Subspace Methods for High-Dimensional Optimization: Sharp Analysis and Lower Bounds

Spectral Norm of Random Kernel Matrices with Applications to Privacy

A Kaczmarz-Inspired Method for Orthogonalization

Generalizing Orthogonalization for Models with Non-Linearities

Learning Random Fourier Features by Hybrid Constrained Optimization

Sharp Matrix Empirical Bernstein Inequalities

Improved Random Features for Dot Product Kernels

Bias-variance decomposition of overparameterized regression with random linear features

Expected number of real zeros for random orthogonal polynomials

Nyström Method vs Random Fourier Features: A Theoretical and Empirical Comparison.

Integrated Variational Fourier Features for Fast Spatial Modelling with Gaussian Processes