Faster Algorithms for Structured Linear and Kernel Support Vector Machines
Organizations: New York University · Simons Institute for the Theory of Computing, UC Berkeley · Massachusetts Institute of Technology
Abstract
Many machine learning problems are quadratic programs, including Support Vector Machines (SVMs). A general quadratic program has input size , and assuming the Strong Exponential Time Hypothesis (SETH), no time algorithm exists when the objective matrix is positive semidefinite (Backurs, Indyk, and Schmidt, NeurIPS'17). SVMs, however, are given by data points of dimension , often much smaller than , and have only equality constraints. In this work, we design the first nearly-linear time algorithm for solving quadratic programs whenever a low-rank factorization of the quadratic objective matrix is given, the number of linear constraints is small, the coefficients of the objective and some optimal solution are bounded by , and some feasible point has all entries between and . For linear SVM on -dimensional data, our randomized algorithm runs in expected time when and the data points have norms at most and margin at least , where is the matrix multiplication exponent. Its output satisfies the equality constraint only approximately. For Gaussian kernel SVM, when , the squared dataset radius is and the dual variables lie in a box with , our randomized algorithm runs in expected time for any accuracy . We also prove that, assuming SETH, when the squared dataset radius is , time is required on some inputs to reach error for the program without the bias term. This improves upon the prior best lower bound in both the dimension and the squared dataset radius.
Figures & tables
Support Vector Machines
Support vector machines learn margin-based classifiers and regressors using linear or kernel-defined decision boundaries.
See all 47 papers