Faster SVM training via conjugate SMO

作者：

Highlights：

• SMO is the state-of-the-art algorithm to solve kernel SVMs.

• We propose CSMO, a simple conjugate descent variant of SMO.

• We prove general convergence and a linear one for definite positive kernel matrices.

• We implement CSMO inside the LIBSVM library, freely available on GitHub.

• Time comparison in 12 large datasets often shows better CSMO performance.

摘要

•SMO is the state-of-the-art algorithm to solve kernel SVMs.•We propose CSMO, a simple conjugate descent variant of SMO.•We prove general convergence and a linear one for definite positive kernel matrices.•We implement CSMO inside the LIBSVM library, freely available on GitHub.•Time comparison in 12 large datasets often shows better CSMO performance.

论文关键词：SVM,Conjugate gradient,SMO

论文评审过程：Received 15 April 2020, Revised 29 July 2020, Accepted 6 September 2020, Available online 17 September 2020, Version of Record 22 September 2020.

论文官网地址：https://doi.org/10.1016/j.patcog.2020.107644