Divide-and-Conquer kernel SVM (DC-SVM)

The Program

DC-SVM implements a divide-and-conquer procedure for speeding up kernel SVM training.
Details can be found in the following paper.

A Divide-and-Conquer Solver for Kernel Support Vector Machines
Cho-Jui Hsieh, Si Si, and Inderjit S. Dhillon
ICML, 2014

Please acknowledge the use of the code with a citation.

Download

You can download the code here. The latest implementation is based on LIBSVM.

Example on ijcnn1 dataset (included in the package)

>> demo_ijcnn
Start training Gaussian kernel SVM with early prediction
RBF kernel, DCSVM-early test accuracy 0.983959, training time 5.57 seconds
Start training Gaussian kernel SVM
Training Level 4
Training Level 3
Training Level 2
Training Level 1
RBF kernel, DC-SVM test accuracy 0.983915, training time 31.44 seconds
Start training polynomial kernel SVM with early prediction
polynomial kernel, DC-SVM early test accuracy 0.985998, training time 29.59 seconds
Start training polynomial kernel SVM
Training Level 4
Training Level 3
Training Level 2
Training Level 1
polynomial kernel, DC-SVM test accuracy 0.971189, training time 43.45 seconds

Dependency