US 11,704,584 B2
Fast and accurate machine learning by applying efficient preconditioner to kernel ridge regression
Gil Shabat, Hod Hasharon (IL); Era Choshen, Pardes Hanna (IL); Dvir Ben-Or, Kfar Saba (IL); and Nadav Carmel, Ramat Gan (IL)
Assigned to Playtika Ltd., Hertsliya (IL)
Filed by Playtika Ltd., Hertsliya (IL)
Filed on May 22, 2020, as Appl. No. 16/881,278.
Prior Publication US 2021/0365820 A1, Nov. 25, 2021
Int. Cl. G06N 20/00 (2019.01); G06N 20/10 (2019.01); G06N 7/00 (2023.01); G06N 5/022 (2023.01)
CPC G06N 7/005 (2013.01) [G06N 5/022 (2013.01); G06N 20/00 (2019.01)] 20 Claims
OG exemplary drawing
 
1. A method for training a machine learning model using Kernel Ridge Regression (KRR), the method comprising:
selecting a plurality of anchor points that represents a subset of columns and/or rows of an initial kernel of a training dataset, without generating the initial kernel, wherein the anchor points are selected by:
generating a randomized decomposition of the initial kernel by projecting the initial kernel onto a random matrix in a lower dimensional space;
permuting the randomized decomposition to reorder the columns and/or rows to approximate the initial kernel; and
selecting anchor points representing a subset of columns and/or rows based on their permuted order;
generating a reduced-rank kernel comprising the subset of columns and/or rows represented by the selected anchor points, wherein the reduced-rank kernel approximates the initial kernel with a relatively lower rank than the initial kernel;
preconditioning a KRR system using a preconditioner generated based on the reduced-rank kernel; and
solving the preconditioned KRR system to train the machine learning model.