Tuesday, 6 June 2017

New Pattern Recognition paper on Indefinite Core Vector Machine


Our new article Indefinite Core Vector Machine is now online at Pattern Recognition (Elsevier)
http://www.sciencedirect.com/science/article/pii/S0031320317302261

Highlights

  • Indefinite Core Vector Machine (iCVM) is proposed
  • approximation concepts are provided leading to linear runtime complexity under moderate assumptions
  • sparsification of iCVM is proposed showing that in many cases also a low memory complexity can be obtained with an acceptable loss in accuracy
  • the algorithm is compared to a number of related methods and multiple datasets showing competitive performance but with much lower computational and memory complexity


Sunday, 28 May 2017

Accepted paper @ ICANN 2017

New paper proposing indefinite Support Vector Regression will be presented at ICANN 2017

Saturday, 31 October 2015

Accepted paper@Simbad 2015

Got a paper accepted about Large scale Indefinite Kernel Fisher Discriminant
to be presented at the next Simbad Workshop. We present a way to get linear
runtime complexity for the iKFD algorithm given the input matrix is (approximately)
low rank. The original iKFD has cubic runtime complexity. iKFD is a very good
classification algorithm for indefinite / non-metric / non-positive input kernels and
our proposal makes iKFD ready for large scale problems.

Friday, 29 May 2015

Review paper on Indefinite proximity learning @ NECO accepted

The paper Indefinite proximity learning - A review by
Frank-Michael Schleif and Peter Tino was accepted for publication
in Neural Computation - the paper is available as open access
(just click on the paper title)

Tuesday, 12 May 2015

New article about conversion of proximity data published

Our article

Metric and non-metric proximity transformations at linear costs

is now online available at Elsevier / Neurocomputing.

Highlights

We propose a linear time and memory efficient approach for converting low rank dissimilarity matrices to similarity matrices and vice versa.
Our approach is applicable for proximities obtained from non-metric proximity measures (indefinite kernels, non-standard dissimilarity measures).
The presented approach also comprises a generalization of Landmark MDS – the presented approach is in general more accurate and flexible than Landmark MDS.
We provide an alternative derivation of the Nyström approximation together with a convergence proof, also for indefinite kernels not given in the workshop paper as a core element of the approach.

Wednesday, 29 April 2015

PCVM library code announced at MLOSS

The pcvm library code has been announced at mloss.org
https://mloss.org/software/view/610/

PCVM C++/armadillo library 0.2

New version of the PCVM C++ code is now available including
the usage of Nystroem approximated kernel (see Readme).
The code now also contains concepts published at ESANN 2015 see
http://www.promos-science.blogspot.de/2015/01/accepted-paper-at-esann2015.html

It may also be interesting if you would like to have a C++/Armadillo implementation
of a Nystroem approximated singular value decomposition (SVD), eigenvalue decomposition (EVD)
or pseudo inverse calculation (PINV). Further it contains code for the normCDF and normPDF.

The code is available at PCVM code