UniverSVM 1.22

Fabian Sinz, Matteo Roffilli — Tue, 16 Oct 2012 11:24:12 -0000

The UniverSVM is a SVM implementation written in C/C++. Its functionality comprises large scale transduction via CCCP optimization, sparse solutions via CCCP optimization and data-dependent regularization with a Universum.

The CCCP optimization procedure allows to train very large transductive SVMs (up to at least 50k unlabeled examples have been tested) or obtain sparse standard SVM solutions that speed up the testing on new examples. In CCCP optimization the loss function is split into a convex and a concave part. In an iterative optimization, the concave part is approximated by its tangent at each step. In order to make the iterative procedure efficient, kernel values are stored in a cache, which is kept for training subsequent SVMs during the iterative optimization.

The data-dependent regularization with a Universum allows to specify implicit invariance directions via an additional dataset, called "The Universum".

The Universum takes data in libsvm format. The unlabeled and/or Universum data can either be given in an extra file or be marked by special reserved labels. Data files can also be specified as an index list into another data file. These index lists also allow relabeling. In this way, only one copy of a possibly large dataset must be kept.

UniverSVM uses the SVM optimizer of Lush written by Leon Bottou and part of the data loading routines of libsvm.

mloss.org UniverSVM

UniverSVM 1.22